From ceefc8ddaa3af6ed21e65b9769f8831af92176c6 Mon Sep 17 00:00:00 2001
From: Parcollet Titouan <parcollet.titouan@gmail.com>
Date: Sun, 11 Feb 2024 21:52:39 +0000
Subject: [PATCH] Shorten the data augmentation in YAML + cleaning (code from
Samsung AI Cambridge) (#2399)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
* shorter augmentations in yaml
* layout to 80 char
* listed label replication
* listed label replication
* listed label replication
* Refact CTC
* Refact transducer
* Refact seq2seq
* call replicate label instead of duplication
* refactor aishell
* refactor aishell
* CommonLanuageÃ
* fix error + CV CTC
* Giga OOF
* Giga OOF
* Giga OOF
* Giga OOF
* Giga OOF
* Giga OOF
* Giga OOF
* Giga OOF
* Giga OOF
* Finishing OOF
* final touch LULZ
* fix tests
* Tests???Ã
* fix augment in some recipes
---------
Co-authored-by: Titouan Parcollet/Embedded AI /SRUK/Engineer/Samsung Electronics <t.parcollet@sruk-ccn4.eu.corp.samsungelectronics.net>
Co-authored-by: Mirco Ravanelli <mirco.ravanelli@gmail.com>
---
.../ASR/CTC/hparams/train_with_wav2vec.yaml | 57 ++----
.../AISHELL-1/ASR/CTC/train_with_wav2vec.py | 2 +
.../AISHELL-1/ASR/seq2seq/hparams/train.yaml | 59 +++---
recipes/AISHELL-1/ASR/seq2seq/train.py | 18 +-
.../hparams/train_ASR_transformer.yaml | 67 ++-----
.../train_ASR_transformer_with_wav2vect.yaml | 47 ++---
recipes/AISHELL-1/ASR/transformer/train.py | 34 ++--
.../ASR/transformer/train_with_wav2vect.py | 18 +-
.../Tokenizer/hparams/tokenizer_bpe5000.yaml | 2 +-
.../train_transformer_tokenizer_bpe5000.yaml | 2 +-
.../hparams/sepformer-aishell1mix2-wham.yaml | 2 +-
.../hparams/sepformer-aishell1mix2.yaml | 2 +-
.../hparams/sepformer-aishell1mix3-wham.yaml | 2 +-
.../hparams/sepformer-aishell1mix3.yaml | 2 +-
.../separation/hparams/convtasnet-cross.yaml | 2 +-
.../hparams/convtasnet-independent.yaml | 2 +-
.../hparams/convtasnet-parallel-noise.yaml | 2 +-
.../hparams/convtasnet-parallel-reverb.yaml | 2 +-
.../hparams/convtasnet-parallel.yaml | 2 +-
recipes/CVSS/S2ST/hparams/train_fr-en.yaml | 2 +-
.../lang_id/hparams/train_ecapa_tdnn.yaml | 21 +--
.../CTC/hparams/train_ar_with_wav2vec.yaml | 46 ++---
.../CTC/hparams/train_de_with_wav2vec.yaml | 46 ++---
.../CTC/hparams/train_en_with_wav2vec.yaml | 46 ++---
.../CTC/hparams/train_es_with_wav2vec.yaml | 46 ++---
.../CTC/hparams/train_fr_with_wav2vec.yaml | 46 ++---
.../CTC/hparams/train_it_with_wav2vec.yaml | 47 ++---
.../CTC/hparams/train_pt_with_wav2vec.yaml | 46 ++---
.../CTC/hparams/train_rw_with_wav2vec.yaml | 46 ++---
.../CTC/hparams/train_zh-CN_with_wav2vec.yaml | 46 ++---
.../ASR/seq2seq/hparams/train_de.yaml | 49 ++---
.../ASR/seq2seq/hparams/train_en.yaml | 50 ++---
.../ASR/seq2seq/hparams/train_es.yaml | 50 ++---
.../ASR/seq2seq/hparams/train_fr.yaml | 50 ++---
.../ASR/seq2seq/hparams/train_it.yaml | 50 ++---
.../ASR/seq2seq/hparams/train_rw.yaml | 50 ++---
.../ASR/transducer/hparams/train_de.yaml | 48 ++---
.../ASR/transducer/hparams/train_fr.yaml | 48 ++---
.../ASR/transducer/hparams/train_it.yaml | 48 ++---
recipes/CommonVoice/ASR/transducer/train.py | 32 ++--
.../hparams/train_ar_hf_whisper.yaml | 44 ++---
.../ASR/transformer/hparams/train_de.yaml | 47 ++---
.../hparams/train_fa_hf_whisper.yaml | 45 ++---
.../ASR/transformer/hparams/train_fr.yaml | 47 ++---
.../hparams/train_fr_hf_whisper.yaml | 45 ++---
.../hparams/train_hi_hf_whisper.yaml | 45 ++---
.../ASR/transformer/hparams/train_it.yaml | 47 ++---
.../hparams/train_it_hf_whisper.yaml | 45 ++---
.../hparams/train_mn_hf_whisper.yaml | 45 ++---
.../hparams/train_sr_hf_whisper.yaml | 45 ++---
recipes/CommonVoice/ASR/transformer/train.py | 34 ++--
.../wav2vec2/hparams/wav2vec2_base.yaml | 10 +-
.../hparams/sepformer-dns-16k.yaml | 2 +-
.../CTC/hparams/train_amh_with_wav2vec.yaml | 43 ++---
.../CTC/hparams/train_dar_with_wav2vec.yaml | 44 ++---
.../CTC/hparams/train_fon_with_wav2vec.yaml | 44 ++---
.../CTC/hparams/train_multi_with_wav2vec.yaml | 44 ++---
.../CTC/hparams/train_sw_with_wav2vec.yaml | 44 ++---
.../CTC/hparams/train_wol_with_wav2vec.yaml | 44 ++---
.../hparams/cnn14_classifier.yaml | 2 +-
.../hparams/conv2d_classifier.yaml | 2 +-
.../ESC50/interpret/hparams/l2i_cnn14.yaml | 2 +-
.../hparams/l2i_conv2dclassifier.yaml | 2 +-
recipes/ESC50/interpret/hparams/nmf.yaml | 2 +-
recipes/ESC50/interpret/hparams/piq.yaml | 2 +-
.../ST/transformer/hparams/conformer.yaml | 2 +-
.../ST/transformer/hparams/transformer.yaml | 2 +-
.../Google-speech-commands/hparams/xvect.yaml | 2 +-
.../hparams/xvect_leaf.yaml | 2 +-
.../hparams/train_with_wav2vec2.yaml | 4 +-
.../AST/transformer/hparams/train_samu.yaml | 2 +-
.../hparams/train_samu_mbart_st.yaml | 2 +-
.../hparams/train_samu_nllb_st.yaml | 2 +-
.../hparams/train_w2v2_mbart_st.yaml | 2 +-
.../hparams/train_w2v2_nllb_st.yaml | 2 +-
.../transformer/hparams/train_w2v2_st.yaml | 2 +-
.../transformer/hparams/conformer_medium.yaml | 4 +-
.../KsponSpeech/LM/hparams/transformer.yaml | 2 +-
.../hparams/5K_unigram_subword_bpe.yaml | 2 +-
.../hparams/sepformer-libri2mix.yaml | 2 +-
.../hparams/sepformer-libri3mix.yaml | 2 +-
recipes/LibriParty/VAD/hparams/train.yaml | 4 +-
.../train_hf_wavlm_average_downsampling.yaml | 122 ++++++------
.../train_hf_wavlm_conv_downsampling.yaml | 122 ++++++------
.../train_hf_wavlm_signal_downsampling.yaml | 119 ++++++------
.../ASR/CTC/hparams/train_hf_wav2vec.yaml | 125 ++++++-------
.../train_hf_wav2vec_rnn_rescoring.yaml | 174 +++++++++---------
...rain_hf_wav2vec_transformer_rescoring.yaml | 136 +++++++-------
.../CTC/hparams/train_hf_whisper_encoder.yaml | 118 ++++++------
.../ASR/CTC/hparams/train_sb_wav2vec.yaml | 119 ++++++------
.../LibriSpeech/ASR/CTC/train_with_wav2vec.py | 11 +-
.../LibriSpeech/ASR/CTC/train_with_whisper.py | 11 +-
.../ASR/seq2seq/hparams/train_BPE_1000.yaml | 128 ++++++-------
.../hparams/train_BPE_1000_sligru.yaml | 129 ++++++-------
.../ASR/seq2seq/hparams/train_BPE_5000.yaml | 129 ++++++-------
recipes/LibriSpeech/ASR/seq2seq/train.py | 14 +-
.../hparams/conformer_transducer.yaml | 60 +++---
recipes/LibriSpeech/ASR/transducer/train.py | 29 +--
.../ASR/transformer/hparams/bayesspeech.yaml | 61 +++---
.../hparams/branchformer_large.yaml | 67 +++----
.../transformer/hparams/conformer_large.yaml | 64 +++----
.../transformer/hparams/conformer_small.yaml | 67 +++----
.../hparams/hyperbranchformer_13M.yaml | 64 +++----
.../hparams/hyperbranchformer_25M.yaml | 64 +++----
.../hparams/hyperconformer_22M.yaml | 65 +++----
.../hparams/hyperconformer_8M.yaml | 64 +++----
.../transformer/hparams/train_hf_whisper.yaml | 53 +++---
.../ASR/transformer/hparams/transformer.yaml | 64 +++----
recipes/LibriSpeech/ASR/transformer/train.py | 18 +-
.../G2P/hparams/hparams_g2p_rnn.yaml | 2 +-
.../G2P/hparams/hparams_g2p_transformer.yaml | 2 +-
.../G2P/hparams/hparams_lm_rnn.yaml | 4 +-
.../G2P/hparams/hparams_lm_transformer.yaml | 4 +-
recipes/LibriSpeech/LM/hparams/RNNLM.yaml | 4 +-
.../LibriSpeech/LM/hparams/transformer.yaml | 2 +-
.../hparams/1K_unigram_subword_bpe.yaml | 2 +-
.../hparams/5K_unigram_subword_bpe.yaml | 2 +-
.../wav2vec2/hparams/wav2vec2_base.yaml | 4 +-
.../ASR/CTC/hparams/train_hf_wav2vec.yaml | 4 +-
.../CTC/hparams/train_hf_wav2vec_full.yaml | 4 +-
.../CTC/hparams/train_hf_wav2vec_relax.yaml | 4 +-
.../gpt/hparams/train_gpt.yaml | 2 +-
.../llama2/hparams/train_llama2.yaml | 2 +-
.../hparams/pool_sisnrestimator.yaml | 2 +-
.../noise-robust/hparams/robust_asr_16k.yaml | 4 +-
recipes/SLURP/NLU/hparams/train.yaml | 4 +-
.../Tokenizer/hparams/tokenizer_bpe58.yaml | 2 +-
recipes/SLURP/direct/hparams/train.yaml | 41 ++---
.../direct/hparams/train_with_wav2vec2.yaml | 42 ++---
.../ASR/CTC/hparams/train_with_wav2vec.yaml | 44 ++---
.../ASR/seq2seq/hparams/train_BPE_2000.yaml | 51 ++---
recipes/Switchboard/ASR/seq2seq/train.py | 14 +-
.../ASR/transformer/hparams/transformer.yaml | 59 ++----
.../hparams/transformer_finetuned_LM.yaml | 53 ++----
recipes/Switchboard/ASR/transformer/train.py | 18 +-
.../Switchboard/LM/hparams/transformer.yaml | 2 +-
.../LM/hparams/transformer_finetune.yaml | 2 +-
.../hparams/2K_unigram_subword_bpe.yaml | 2 +-
recipes/TIMIT/ASR/CTC/hparams/train.yaml | 52 ++----
recipes/TIMIT/ASR/seq2seq/hparams/train.yaml | 45 ++---
.../seq2seq/hparams/train_with_wav2vec2.yaml | 44 ++---
.../TIMIT/ASR/transducer/hparams/train.yaml | 52 ++----
.../ASR/transducer/hparams/train_wav2vec.yaml | 45 ++---
recipes/TIMIT/Alignment/hparams/train.yaml | 4 +-
.../hparams/branchformer_large.yaml | 4 +-
.../Tokenizer/hparams/tedlium2_500_bpe.yaml | 2 +-
.../hparams/train_ecapa_tdnn.yaml | 2 +-
recipes/Voicebank/ASR/CTC/hparams/train.yaml | 44 ++---
.../ASR_enhance/hparams/enhance_mimic.yaml | 2 +-
.../hparams/pretrain_perceptual.yaml | 2 +-
.../MTL/ASR_enhance/hparams/robust_asr.yaml | 49 ++---
.../SpeakerRec/hparams/train_ecapa_tdnn.yaml | 40 ++--
.../hparams/train_ecapa_tdnn_mel_spec.yaml | 40 ++--
.../SpeakerRec/hparams/train_resnet.yaml | 42 ++---
.../SpeakerRec/hparams/train_x_vectors.yaml | 40 ++--
.../lang_id/hparams/train_ecapa.yaml | 13 +-
.../hparams/cnntransformer-wham-DM.yaml | 2 +-
.../hparams/cnntransformer-whamr-DM.yaml | 2 +-
.../hparams/convtasnet-whamr-DM.yaml | 2 +-
.../enhancement/hparams/dprnn-whamr-DM.yaml | 2 +-
.../enhancement/hparams/sepformer-wham.yaml | 2 +-
.../hparams/sepformer-whamr-16k-DM.yaml | 2 +-
.../hparams/sepformer-whamr-16k.yaml | 2 +-
.../hparams/sepformer-whamr-DM.yaml | 2 +-
.../enhancement/hparams/sepformer-whamr.yaml | 2 +-
.../separation/hparams/sepformer-wham.yaml | 2 +-
.../separation/hparams/sepformer-whamr.yaml | 2 +-
.../separation/hparams/convtasnet.yaml | 2 +-
recipes/WSJ0Mix/separation/hparams/dprnn.yaml | 2 +-
.../separation/hparams/resepformer.yaml | 2 +-
.../hparams/sepformer-conformerintra.yaml | 2 +-
.../hparams/sepformer-customdataset.yaml | 2 +-
.../WSJ0Mix/separation/hparams/sepformer.yaml | 2 +-
recipes/WSJ0Mix/separation/hparams/skim.yaml | 2 +-
.../emotion_diarization/hparams/train.yaml | 2 +-
.../Tokenizer/hparams/tokenizer_bpe51.yaml | 2 +-
.../direct/hparams/train.yaml | 44 ++---
recipes/timers-and-such/LM/hparams/train.yaml | 2 +-
.../Tokenizer/hparams/tokenizer_bpe51.yaml | 2 +-
.../decoupled/hparams/train_LS_LM.yaml | 2 +-
.../decoupled/hparams/train_TAS_LM.yaml | 2 +-
.../timers-and-such/direct/hparams/train.yaml | 50 ++---
.../direct/hparams/train_with_wav2vec2.yaml | 42 ++---
.../multistage/hparams/train_LS_LM.yaml | 51 ++---
.../multistage/hparams/train_TAS_LM.yaml | 50 ++---
speechbrain/augment/augmenter.py | 39 +++-
186 files changed, 2124 insertions(+), 3354 deletions(-)
diff --git a/recipes/AISHELL-1/ASR/CTC/hparams/train_with_wav2vec.yaml b/recipes/AISHELL-1/ASR/CTC/hparams/train_with_wav2vec.yaml
index 9bd0b52d2..486685f25 100644
--- a/recipes/AISHELL-1/ASR/CTC/hparams/train_with_wav2vec.yaml
+++ b/recipes/AISHELL-1/ASR/CTC/hparams/train_with_wav2vec.yaml
@@ -28,7 +28,8 @@ test_data: !ref <output_folder>/test.csv
wav2vec2_hub: TencentGameMate/chinese-wav2vec2-large
wav2vec2_folder: !ref <save_folder>/wav2vec2_checkpoint
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 80
lr: 1.0
lr_wav2vec: 0.0001
@@ -76,7 +77,8 @@ tokenizer: !apply:transformers.BertTokenizer.from_pretrained
# bert-base-chinese tokens length
output_neurons: 21128
-# Decoding parameters
+############################## Decoding ########################################
+
# Be sure that the bos and eos index match with the BPEs ones
# Decoding parameters
test_searcher: !name:speechbrain.decoders.CTCBeamSearcher
@@ -98,64 +100,37 @@ beta: 0.5
# which Chinese writing normally does not do.
# If remove_spaces, spaces are removed
# from the transcript before computing CER.
-# (e.g., 祝 可爱 的 你 —> 祝可爱的你)
remove_spaces: True
split_tokens: !apply:operator.not_ [!ref <remove_spaces>]
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [90, 100, 110] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
# Time Drop
-time_drop_length_low: 35 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 45 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 2 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 2 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
-
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 35
+ drop_length_high: 45
+ drop_count_low: 2
+ drop_count_high: 2
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 2
+ drop_count_high: 2
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -164,6 +139,8 @@ fea_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <freq_drop>,
!ref <time_warp>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
linear1: !name:speechbrain.nnet.linear.Linear
@@ -230,6 +207,8 @@ lr_annealing_wav2vec: !new:speechbrain.nnet.schedulers.NewBobScheduler
annealing_factor: 0.9
patient: 0
+############################## Logging and Pretrainer ##########################
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
recoverables:
diff --git a/recipes/AISHELL-1/ASR/CTC/train_with_wav2vec.py b/recipes/AISHELL-1/ASR/CTC/train_with_wav2vec.py
index 227204f44..43783eed7 100644
--- a/recipes/AISHELL-1/ASR/CTC/train_with_wav2vec.py
+++ b/recipes/AISHELL-1/ASR/CTC/train_with_wav2vec.py
@@ -56,6 +56,8 @@ class ASR(sb.Brain):
ids = batch.id
tokens, tokens_lens = batch.tokens
+ # Labels must be extended if parallel augmentation or concatenated
+ # augmentation was performed on the input (increasing the time dimension)
if stage == sb.Stage.TRAIN:
if hasattr(self.hparams, "fea_augment"):
tokens = self.hparams.fea_augment.replicate_labels(tokens)
diff --git a/recipes/AISHELL-1/ASR/seq2seq/hparams/train.yaml b/recipes/AISHELL-1/ASR/seq2seq/hparams/train.yaml
index 75b303f66..e6fda7de2 100644
--- a/recipes/AISHELL-1/ASR/seq2seq/hparams/train.yaml
+++ b/recipes/AISHELL-1/ASR/seq2seq/hparams/train.yaml
@@ -29,7 +29,8 @@ test_data: !ref <output_folder>/test.csv
noise_annotation: !ref <save_folder>/noise.csv #The data manifest files are created by the data preparation script
tokenizer_file: speechbrain/asr-transformer-aishell/tokenizer.ckpt
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 40
number_of_ctc_epochs: 10
batch_size: 16
@@ -71,7 +72,7 @@ test_dataloader_opts:
batch_size: !ref <batch_size>
num_workers: !ref <num_workers>
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 2
@@ -108,7 +109,6 @@ scorer_beam_scale: 0.5
# which Chinese writing normally does not do.
# If remove_spaces, spaces are removed
# from the transcript before computing CER.
-# (e.g., 祝 可爱 的 你 —> 祝可爱的你)
remove_spaces: True
split_tokens: !apply:operator.not_ [!ref <remove_spaces>]
@@ -118,6 +118,8 @@ epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
+############################## Augmentations ###################################
+
compute_features: !new:speechbrain.lobes.features.Fbank
sample_rate: !ref <sample_rate>
n_fft: !ref <n_fft>
@@ -132,57 +134,37 @@ prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_U
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
-# Augmenter: Combines previously defined augmentations to perform data augmentation
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -192,6 +174,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.lobes.models.CRDNN.CRDNN
input_shape: [null, null, !ref <n_mels>]
activation: !ref <activation>
@@ -268,7 +252,8 @@ pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
paths:
tokenizer: !ref <tokenizer_file>
-# Scorer
+############################## Decoding ########################################
+
ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
eos_index: !ref <eos_index>
blank_index: !ref <blank_index>
@@ -305,6 +290,8 @@ lr_annealing: !new:speechbrain.nnet.schedulers.NewBobScheduler
annealing_factor: 0.8
patient: 0
+############################## Logging and Pretrainer ##########################
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
recoverables:
diff --git a/recipes/AISHELL-1/ASR/seq2seq/train.py b/recipes/AISHELL-1/ASR/seq2seq/train.py
index 69d2e75d6..bc2c49b88 100644
--- a/recipes/AISHELL-1/ASR/seq2seq/train.py
+++ b/recipes/AISHELL-1/ASR/seq2seq/train.py
@@ -29,10 +29,6 @@ class ASR(sb.Brain):
# Forward pass
feats = self.hparams.compute_features(wavs)
-
- if stage == sb.Stage.TRAIN and hasattr(self.hparams, "fea_augment"):
- feats, fea_lens = self.hparams.fea_augment(feats, wav_lens)
-
feats = self.modules.normalize(feats, wav_lens)
x = self.modules.enc(feats.detach())
e_in = self.modules.emb(tokens_bos) # y_in bos + tokens
@@ -65,12 +61,16 @@ class ASR(sb.Brain):
tokens_eos, tokens_eos_lens = batch.tokens_eos
tokens, tokens_lens = batch.tokens
+ # Labels must be extended if parallel augmentation or concatenated
+ # augmentation was performed on the input (increasing the time dimension)
if stage == sb.Stage.TRAIN and hasattr(self.hparams, "wav_augment"):
- tokens = self.hparams.wav_augment.replicate_labels(tokens)
- tokens_lens = self.hparams.wav_augment.replicate_labels(tokens_lens)
- tokens_eos = self.hparams.wav_augment.replicate_labels(tokens_eos)
- tokens_eos_lens = self.hparams.wav_augment.replicate_labels(
- tokens_eos_lens
+ (
+ tokens,
+ tokens_lens,
+ tokens_eos,
+ tokens_eos_lens,
+ ) = self.hparams.wav_augment.replicate_multiple_labels(
+ tokens, tokens_lens, tokens_eos, tokens_eos_lens
)
loss_seq = self.hparams.seq_cost(
diff --git a/recipes/AISHELL-1/ASR/transformer/hparams/train_ASR_transformer.yaml b/recipes/AISHELL-1/ASR/transformer/hparams/train_ASR_transformer.yaml
index b98e371b2..408c9e680 100644
--- a/recipes/AISHELL-1/ASR/transformer/hparams/train_ASR_transformer.yaml
+++ b/recipes/AISHELL-1/ASR/transformer/hparams/train_ASR_transformer.yaml
@@ -30,7 +30,8 @@ test_data: !ref <save_folder>/test.csv
noise_annotation: !ref <save_folder>/noise.csv #The data manifest files are created by the data preparation script
tokenizer_file: speechbrain/asr-transformer-aishell/tokenizer.ckpt
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 50
batch_size: 8
ctc_weight: 0.3
@@ -77,7 +78,7 @@ test_dataloader_opts:
batch_size: !ref <batch_size>
num_workers: !ref <num_workers>
-####################### Model parameters ###########################
+####################### Model Parameters #######################################
# Transformer
d_model: 256
nhead: 4
@@ -103,7 +104,7 @@ valid_beam_size: 10
test_beam_size: 10
ctc_weight_decode: 0.40
-############################## models ################################
+############################## Models ##########################################
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
input_shape: (8, 10, 80)
@@ -157,7 +158,8 @@ SGD: !name:torch.optim.SGD
momentum: 0.99
nesterov: True
-# Scorer
+############################## Decoding & optimiser ############################
+
ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
eos_index: !ref <eos_index>
blank_index: !ref <blank_index>
@@ -227,7 +229,7 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
-# ----- WAVEFORM AUGMENTATION ----- #
+############################## Augmentation ####################################
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
@@ -238,75 +240,43 @@ prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_U
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 1
max_augmentations: 1
augment_prob: 1.0
augmentations: [
!ref <add_noise>]
-
- # ----- FEATURE AUGMENTATION ----- #
-time_drop_length_low: 0 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 100 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 2 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 2 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
-
# Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
-
-freq_drop_length_low: 30 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 40 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
+ drop_length_low: 0
+ drop_length_high: 100
+ drop_count_low: 2
+ drop_count_high: 2
# Frequency Drop
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 30
+ drop_length_high: 40
+ drop_count_low: 2
+ drop_count_high: 2
dim: 2
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
# Time warp
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 1
max_augmentations: 1
augment_start_index: !ref <batch_size> # This leaves unchanges original inputs
@@ -317,6 +287,8 @@ fea_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <freq_drop>,
!ref <time_warp>]
+############################## Logging and Pretrainer ##########################
+
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
save_file: !ref <train_log>
@@ -324,7 +296,6 @@ train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
# which Chinese writing normally does not do.
# If remove_spaces, spaces are removed
# from the transcript before computing CER.
-# (e.g., 祝 可爱 的 你 —> 祝可爱的你)
remove_spaces: True
split_tokens: !apply:operator.not_ [!ref <remove_spaces>]
diff --git a/recipes/AISHELL-1/ASR/transformer/hparams/train_ASR_transformer_with_wav2vect.yaml b/recipes/AISHELL-1/ASR/transformer/hparams/train_ASR_transformer_with_wav2vect.yaml
index 13a7826ad..a196afc58 100644
--- a/recipes/AISHELL-1/ASR/transformer/hparams/train_ASR_transformer_with_wav2vect.yaml
+++ b/recipes/AISHELL-1/ASR/transformer/hparams/train_ASR_transformer_with_wav2vect.yaml
@@ -30,7 +30,8 @@ wav2vec2_hub: facebook/wav2vec2-large-100k-voxpopuli
wav2vec2_folder: !ref <save_folder>/wav2vec2_checkpoint
freeze_wav2vec: False
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 80
batch_size: 2
grad_accumulation_factor: 16
@@ -72,7 +73,7 @@ test_dataloader_opts:
batch_size: !ref <batch_size>
num_workers: !ref <num_workers>
-####################### Model parameters ###########################
+####################### Model Parameters #######################################
# Transformer
d_model: 256
nhead: 4
@@ -98,7 +99,7 @@ valid_beam_size: 10
test_beam_size: 10
ctc_weight_decode: 0.40
-############################## models ################################
+############################## Models ##########################################
wav2vec2: !new:speechbrain.lobes.models.huggingface_transformers.wav2vec2.Wav2Vec2
source: !ref <wav2vec2_hub>
@@ -140,44 +141,27 @@ model: !new:torch.nn.ModuleList
- [!ref <Transformer>, !ref <seq_lin>, !ref <ctc_lin>]
# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -186,6 +170,7 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Decoding & optimiser ############################
# define two optimizers here for two-stage training
Adam: !name:torch.optim.Adam
@@ -257,6 +242,7 @@ noam_annealing_wav2vect: !new:speechbrain.nnet.schedulers.NoamScheduler
n_warmup_steps: 25000
model_size: !ref <d_model>
+############################## Logging and Pretrainer ##########################
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
@@ -278,7 +264,6 @@ train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
# which Chinese writing normally does not do.
# If remove_spaces, spaces are removed
# from the transcript before computing CER.
-# (e.g., 祝 可爱 的 你 —> 祝可爱的你)
remove_spaces: True
split_tokens: !apply:operator.not_ [!ref <remove_spaces>]
diff --git a/recipes/AISHELL-1/ASR/transformer/train.py b/recipes/AISHELL-1/ASR/transformer/train.py
index 977563ac8..63361bf0d 100644
--- a/recipes/AISHELL-1/ASR/transformer/train.py
+++ b/recipes/AISHELL-1/ASR/transformer/train.py
@@ -82,28 +82,26 @@ class ASR(sb.core.Brain):
tokens, tokens_lens = batch.tokens
if stage == sb.Stage.TRAIN:
+ # Labels must be extended if parallel augmentation or concatenated
+ # augmentation was performed on the input (increasing the time dimension)
if hasattr(self.hparams, "wav_augment"):
- tokens = self.hparams.wav_augment.replicate_labels(tokens)
- tokens_lens = self.hparams.wav_augment.replicate_labels(
- tokens_lens
- )
- tokens_eos = self.hparams.wav_augment.replicate_labels(
- tokens_eos
- )
- tokens_eos_lens = self.hparams.wav_augment.replicate_labels(
- tokens_eos_lens
+ (
+ tokens,
+ tokens_lens,
+ tokens_eos,
+ tokens_eos_lens,
+ ) = self.hparams.wav_augment.replicate_multiple_labels(
+ tokens, tokens_lens, tokens_eos, tokens_eos_lens
)
if hasattr(self.hparams, "fea_augment"):
- tokens = self.hparams.fea_augment.replicate_labels(tokens)
- tokens_lens = self.hparams.fea_augment.replicate_labels(
- tokens_lens
- )
- tokens_eos = self.hparams.wav_augment.replicate_labels(
- tokens_eos
- )
- tokens_eos_lens = self.hparams.wav_augment.replicate_labels(
- tokens_eos_lens
+ (
+ tokens,
+ tokens_lens,
+ tokens_eos,
+ tokens_eos_lens,
+ ) = self.hparams.fea_augment.replicate_multiple_labels(
+ tokens, tokens_lens, tokens_eos, tokens_eos_lens
)
loss_seq = self.hparams.seq_cost(
diff --git a/recipes/AISHELL-1/ASR/transformer/train_with_wav2vect.py b/recipes/AISHELL-1/ASR/transformer/train_with_wav2vect.py
index 94196ea7b..53aa47375 100644
--- a/recipes/AISHELL-1/ASR/transformer/train_with_wav2vect.py
+++ b/recipes/AISHELL-1/ASR/transformer/train_with_wav2vect.py
@@ -74,16 +74,16 @@ class ASR(sb.core.Brain):
tokens, tokens_lens = batch.tokens
if stage == sb.Stage.TRAIN:
+ # Labels must be extended if parallel augmentation or concatenated
+ # augmentation was performed on the input (increasing the time dimension)
if hasattr(self.hparams, "wav_augment"):
- tokens = self.hparams.wav_augment.replicate_labels(tokens)
- tokens_lens = self.hparams.wav_augment.replicate_labels(
- tokens_lens
- )
- tokens_eos = self.hparams.wav_augment.replicate_labels(
- tokens_eos
- )
- tokens_eos_lens = self.hparams.wav_augment.replicate_labels(
- tokens_eos_lens
+ (
+ tokens,
+ tokens_lens,
+ tokens_eos,
+ tokens_eos_lens,
+ ) = self.hparams.wav_augment.replicate_multiple_labels(
+ tokens, tokens_lens, tokens_eos, tokens_eos_lens
)
loss_seq = self.hparams.seq_cost(
diff --git a/recipes/AISHELL-1/Tokenizer/hparams/tokenizer_bpe5000.yaml b/recipes/AISHELL-1/Tokenizer/hparams/tokenizer_bpe5000.yaml
index 886d22bda..d2cb23018 100644
--- a/recipes/AISHELL-1/Tokenizer/hparams/tokenizer_bpe5000.yaml
+++ b/recipes/AISHELL-1/Tokenizer/hparams/tokenizer_bpe5000.yaml
@@ -15,7 +15,7 @@ train_csv: !ref <output_folder>/train.csv
valid_csv: !ref <output_folder>/dev.csv
-# Training parameters
+####################### Training Parameters ####################################
token_type: unigram # ["unigram", "bpe", "char"]
token_output: 5000 # index(blank/eos/bos/unk) = 0
character_coverage: 1.0
diff --git a/recipes/AISHELL-1/Tokenizer/hparams/train_transformer_tokenizer_bpe5000.yaml b/recipes/AISHELL-1/Tokenizer/hparams/train_transformer_tokenizer_bpe5000.yaml
index bc286156b..973df9a11 100644
--- a/recipes/AISHELL-1/Tokenizer/hparams/train_transformer_tokenizer_bpe5000.yaml
+++ b/recipes/AISHELL-1/Tokenizer/hparams/train_transformer_tokenizer_bpe5000.yaml
@@ -15,7 +15,7 @@ train_csv: !ref <output_folder>/train.csv
valid_csv: !ref <output_folder>/dev.csv
-# Training parameters
+####################### Training Parameters ####################################
token_type: unigram # ["unigram", "bpe", "char"]
token_output: 5000 # index(blank/eos/bos/unk) = 0
character_coverage: 1.0
diff --git a/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix2-wham.yaml b/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix2-wham.yaml
index 24571404f..d3cb9493e 100644
--- a/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix2-wham.yaml
+++ b/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix2-wham.yaml
@@ -40,7 +40,7 @@ noprogressbar: False
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix2.yaml b/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix2.yaml
index d4689378a..168471dbb 100644
--- a/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix2.yaml
+++ b/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix2.yaml
@@ -40,7 +40,7 @@ noprogressbar: False
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix3-wham.yaml b/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix3-wham.yaml
index 65add025a..834857ed7 100644
--- a/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix3-wham.yaml
+++ b/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix3-wham.yaml
@@ -40,7 +40,7 @@ noprogressbar: False
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix3.yaml b/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix3.yaml
index 3b27b796b..d48fdecb2 100644
--- a/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix3.yaml
+++ b/recipes/Aishell1Mix/separation/hparams/sepformer-aishell1mix3.yaml
@@ -40,7 +40,7 @@ noprogressbar: False
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-cross.yaml b/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-cross.yaml
index f609c746e..043845aeb 100644
--- a/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-cross.yaml
+++ b/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-cross.yaml
@@ -43,7 +43,7 @@ save_audio: True # Save estimated sources on disk
n_audio_to_save: 10
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-independent.yaml b/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-independent.yaml
index c3483f5f2..164ccc45b 100644
--- a/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-independent.yaml
+++ b/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-independent.yaml
@@ -43,7 +43,7 @@ save_audio: True # Save estimated sources on disk
n_audio_to_save: 10
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-parallel-noise.yaml b/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-parallel-noise.yaml
index 74941dfea..fef85267f 100644
--- a/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-parallel-noise.yaml
+++ b/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-parallel-noise.yaml
@@ -43,7 +43,7 @@ save_audio: True # Save estimated sources on disk
n_audio_to_save: 10
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-parallel-reverb.yaml b/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-parallel-reverb.yaml
index 6b1518d39..4ec5054f9 100644
--- a/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-parallel-reverb.yaml
+++ b/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-parallel-reverb.yaml
@@ -43,7 +43,7 @@ save_audio: True # Save estimated sources on disk
n_audio_to_save: 10
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-parallel.yaml b/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-parallel.yaml
index a5d615679..adb31ddc6 100644
--- a/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-parallel.yaml
+++ b/recipes/BinauralWSJ0Mix/separation/hparams/convtasnet-parallel.yaml
@@ -43,7 +43,7 @@ save_audio: True # Save estimated sources on disk
n_audio_to_save: 10
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/CVSS/S2ST/hparams/train_fr-en.yaml b/recipes/CVSS/S2ST/hparams/train_fr-en.yaml
index 8ff7d59be..678dd1c17 100644
--- a/recipes/CVSS/S2ST/hparams/train_fr-en.yaml
+++ b/recipes/CVSS/S2ST/hparams/train_fr-en.yaml
@@ -59,7 +59,7 @@ wav2vec2_download_path: !ref <save_folder>/pretrained_models
wav2vec2_frozen: False
wav2vec2_freeze_steps: 10000
-# Training parameters
+####################### Training Parameters ####################################
lr: 0.0005
lr_wav2vec: 0.00001
loss_reduction: batchmean
diff --git a/recipes/CommonLanguage/lang_id/hparams/train_ecapa_tdnn.yaml b/recipes/CommonLanguage/lang_id/hparams/train_ecapa_tdnn.yaml
index 0c4d91aa9..d4722f45c 100644
--- a/recipes/CommonLanguage/lang_id/hparams/train_ecapa_tdnn.yaml
+++ b/recipes/CommonLanguage/lang_id/hparams/train_ecapa_tdnn.yaml
@@ -38,10 +38,10 @@ error_stats: !name:speechbrain.utils.metric_stats.MetricStats
metric: !name:speechbrain.nnet.losses.classification_error
reduction: batch
+####################### Training Parameters ####################################
+
# Feature parameters btw: 40 - 80
n_mels: 80
-
-# Training Parameters
sample_rate: 16000
number_of_epochs: 30
batch_size: 4
@@ -64,6 +64,8 @@ test_dataloader_options:
batch_size: !ref <batch_size>
shuffle: True
+############################## Augmentations ###################################
+
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
URL: !ref <NOISE_DATASET_URL>
@@ -78,7 +80,6 @@ prepare_rir_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
ext: wav
csv_file: !ref <rir_annotation>
-
# Add reverberation to input signal
add_reverb: !new:speechbrain.augment.time_domain.AddReverb
csv_file: !ref <rir_annotation>
@@ -87,27 +88,21 @@ add_reverb: !new:speechbrain.augment.time_domain.AddReverb
num_workers: !ref <num_workers>
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
# Speed perturbation
-speed_changes: [90, 100, 110] # List of speed changes for time-stretching
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
shuffle_augmentations: True
min_augmentations: 1
max_augmentations: 3
@@ -125,6 +120,8 @@ mean_var_norm_input: !new:speechbrain.processing.features.InputNormalization
norm_type: sentence
std_norm: False
+############################## Models ##########################################
+
# To design a custom model, either just edit the simple CustomModel
# class that's listed here, or replace this `!new` call with a line
# pointing to a different file you've defined.
@@ -182,6 +179,8 @@ lr_annealing: !new:speechbrain.nnet.schedulers.LinearScheduler
final_value: !ref <lr_final>
epoch_count: !ref <number_of_epochs>
+############################## Logging and Pretrainer ##########################
+
# This object is used for saving the state of training both so that it
# can be resumed if it gets interrupted, and also so that the best checkpoint
# can be later loaded for evaluation or inference.
diff --git a/recipes/CommonVoice/ASR/CTC/hparams/train_ar_with_wav2vec.yaml b/recipes/CommonVoice/ASR/CTC/hparams/train_ar_with_wav2vec.yaml
index caf7f2d3b..643df0994 100644
--- a/recipes/CommonVoice/ASR/CTC/hparams/train_ar_with_wav2vec.yaml
+++ b/recipes/CommonVoice/ASR/CTC/hparams/train_ar_with_wav2vec.yaml
@@ -32,7 +32,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -58,7 +59,7 @@ test_dataloader_options:
token_type: unigram # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
wav2vec_output_dim: 1024
dnn_neurons: 1024
freeze_wav2vec: False
@@ -97,45 +98,30 @@ test_beam_search:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -144,6 +130,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
linear1: !name:speechbrain.nnet.linear.Linear
diff --git a/recipes/CommonVoice/ASR/CTC/hparams/train_de_with_wav2vec.yaml b/recipes/CommonVoice/ASR/CTC/hparams/train_de_with_wav2vec.yaml
index dc4ae34b2..adb8e5bb5 100644
--- a/recipes/CommonVoice/ASR/CTC/hparams/train_de_with_wav2vec.yaml
+++ b/recipes/CommonVoice/ASR/CTC/hparams/train_de_with_wav2vec.yaml
@@ -33,7 +33,8 @@ skip_prep: False
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 45
lr: 1.0
lr_wav2vec: 0.0001
@@ -61,7 +62,7 @@ test_dataloader_options:
token_type: char # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
# activation: !name:torch.nn.LeakyReLU
dnn_neurons: 1024
wav2vec_output_dim: !ref <dnn_neurons>
@@ -95,45 +96,30 @@ test_beam_search:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -142,6 +128,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
linear1: !name:speechbrain.nnet.linear.Linear
diff --git a/recipes/CommonVoice/ASR/CTC/hparams/train_en_with_wav2vec.yaml b/recipes/CommonVoice/ASR/CTC/hparams/train_en_with_wav2vec.yaml
index f3c68ee9b..d8aaea36e 100644
--- a/recipes/CommonVoice/ASR/CTC/hparams/train_en_with_wav2vec.yaml
+++ b/recipes/CommonVoice/ASR/CTC/hparams/train_en_with_wav2vec.yaml
@@ -32,7 +32,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -58,7 +59,7 @@ test_dataloader_options:
token_type: unigram # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
# activation: !name:torch.nn.LeakyReLU
wav2vec_output_dim: 1024
dnn_neurons: 1024
@@ -95,45 +96,30 @@ test_beam_search:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -142,6 +128,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
linear1: !name:speechbrain.nnet.linear.Linear
diff --git a/recipes/CommonVoice/ASR/CTC/hparams/train_es_with_wav2vec.yaml b/recipes/CommonVoice/ASR/CTC/hparams/train_es_with_wav2vec.yaml
index 8e2056f83..e32a242d1 100644
--- a/recipes/CommonVoice/ASR/CTC/hparams/train_es_with_wav2vec.yaml
+++ b/recipes/CommonVoice/ASR/CTC/hparams/train_es_with_wav2vec.yaml
@@ -32,7 +32,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -58,7 +59,7 @@ test_dataloader_options:
token_type: unigram # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
wav2vec_output_dim: 1024
dnn_neurons: 1024
freeze_wav2vec: False
@@ -96,45 +97,30 @@ test_beam_search:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -143,6 +129,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
linear1: !name:speechbrain.nnet.linear.Linear
diff --git a/recipes/CommonVoice/ASR/CTC/hparams/train_fr_with_wav2vec.yaml b/recipes/CommonVoice/ASR/CTC/hparams/train_fr_with_wav2vec.yaml
index 44f1523f3..079cfe73f 100644
--- a/recipes/CommonVoice/ASR/CTC/hparams/train_fr_with_wav2vec.yaml
+++ b/recipes/CommonVoice/ASR/CTC/hparams/train_fr_with_wav2vec.yaml
@@ -32,7 +32,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -58,7 +59,7 @@ test_dataloader_options:
token_type: char # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
# activation: !name:torch.nn.LeakyReLU
wav2vec_output_dim: 1024
dnn_neurons: 1024
@@ -94,45 +95,30 @@ test_beam_search:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -141,6 +127,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
linear1: !name:speechbrain.nnet.linear.Linear
diff --git a/recipes/CommonVoice/ASR/CTC/hparams/train_it_with_wav2vec.yaml b/recipes/CommonVoice/ASR/CTC/hparams/train_it_with_wav2vec.yaml
index 4f39ad2a0..033299752 100644
--- a/recipes/CommonVoice/ASR/CTC/hparams/train_it_with_wav2vec.yaml
+++ b/recipes/CommonVoice/ASR/CTC/hparams/train_it_with_wav2vec.yaml
@@ -33,7 +33,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 8.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 45
lr: 1.0
lr_wav2vec: 0.0001
@@ -59,7 +60,7 @@ test_dataloader_options:
token_type: unigram # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
# activation: !name:torch.nn.LeakyReLU
wav2vec_output_dim: 1024
dnn_neurons: 1024
@@ -95,45 +96,30 @@ test_beam_search:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -141,6 +127,9 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <speed_perturb>,
!ref <drop_freq>,
!ref <drop_chunk>]
+
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
linear1: !name:speechbrain.nnet.linear.Linear
diff --git a/recipes/CommonVoice/ASR/CTC/hparams/train_pt_with_wav2vec.yaml b/recipes/CommonVoice/ASR/CTC/hparams/train_pt_with_wav2vec.yaml
index da8a28de5..d4b703eb4 100644
--- a/recipes/CommonVoice/ASR/CTC/hparams/train_pt_with_wav2vec.yaml
+++ b/recipes/CommonVoice/ASR/CTC/hparams/train_pt_with_wav2vec.yaml
@@ -31,7 +31,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -57,7 +58,7 @@ test_dataloader_options:
token_type: unigram # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
wav2vec_output_dim: 1024
dnn_neurons: 1024
freeze_wav2vec: False
@@ -95,45 +96,30 @@ test_beam_search:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -142,6 +128,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
linear1: !name:speechbrain.nnet.linear.Linear
diff --git a/recipes/CommonVoice/ASR/CTC/hparams/train_rw_with_wav2vec.yaml b/recipes/CommonVoice/ASR/CTC/hparams/train_rw_with_wav2vec.yaml
index f92d8ad13..ed15a8aad 100644
--- a/recipes/CommonVoice/ASR/CTC/hparams/train_rw_with_wav2vec.yaml
+++ b/recipes/CommonVoice/ASR/CTC/hparams/train_rw_with_wav2vec.yaml
@@ -32,7 +32,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -59,7 +60,7 @@ test_dataloader_options:
token_type: unigram # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
# activation: !name:torch.nn.LeakyReLU
wav2vec_output_dim: 1024
dnn_neurons: 1024
@@ -95,45 +96,30 @@ test_beam_search:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -142,6 +128,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
linear1: !name:speechbrain.nnet.linear.Linear
diff --git a/recipes/CommonVoice/ASR/CTC/hparams/train_zh-CN_with_wav2vec.yaml b/recipes/CommonVoice/ASR/CTC/hparams/train_zh-CN_with_wav2vec.yaml
index 513d8d324..a1709931a 100644
--- a/recipes/CommonVoice/ASR/CTC/hparams/train_zh-CN_with_wav2vec.yaml
+++ b/recipes/CommonVoice/ASR/CTC/hparams/train_zh-CN_with_wav2vec.yaml
@@ -32,7 +32,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -59,7 +60,7 @@ test_dataloader_options:
token_type: unigram # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
wav2vec_output_dim: 1024
dnn_neurons: 1024
freeze_wav2vec: False
@@ -97,45 +98,30 @@ test_beam_search:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -144,6 +130,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
linear1: !name:speechbrain.nnet.linear.Linear
diff --git a/recipes/CommonVoice/ASR/seq2seq/hparams/train_de.yaml b/recipes/CommonVoice/ASR/seq2seq/hparams/train_de.yaml
index b66374147..cb6f2b3be 100644
--- a/recipes/CommonVoice/ASR/seq2seq/hparams/train_de.yaml
+++ b/recipes/CommonVoice/ASR/seq2seq/hparams/train_de.yaml
@@ -30,7 +30,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 25
number_of_ctc_epochs: 20
lr: 1.0
@@ -62,7 +63,7 @@ sample_rate: 16000
n_fft: 400
n_mels: 80
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 3
@@ -104,51 +105,27 @@ temperature: 1.50
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
+############################## Augmentations ###################################
# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
-
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 2
+ drop_count_high: 2
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -165,6 +142,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Models ##########################################
+
enc: !new:speechbrain.lobes.models.CRDNN.CRDNN
input_shape: [null, null, !ref <n_mels>]
activation: !ref <activation>
diff --git a/recipes/CommonVoice/ASR/seq2seq/hparams/train_en.yaml b/recipes/CommonVoice/ASR/seq2seq/hparams/train_en.yaml
index c74d25c66..49f9a0d2b 100644
--- a/recipes/CommonVoice/ASR/seq2seq/hparams/train_en.yaml
+++ b/recipes/CommonVoice/ASR/seq2seq/hparams/train_en.yaml
@@ -29,7 +29,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 25
number_of_ctc_epochs: 10
lr: 1.0
@@ -60,7 +61,7 @@ sample_rate: 16000
n_fft: 400
n_mels: 80
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 3
@@ -102,50 +103,27 @@ temperature: 1.50
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
+############################## Augmentations ###################################
+ # Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 2
+ drop_count_high: 2
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -162,6 +140,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Models ##########################################
+
enc: !new:speechbrain.lobes.models.CRDNN.CRDNN
input_shape: [null, null, !ref <n_mels>]
activation: !ref <activation>
diff --git a/recipes/CommonVoice/ASR/seq2seq/hparams/train_es.yaml b/recipes/CommonVoice/ASR/seq2seq/hparams/train_es.yaml
index b56a75b69..b94373e9b 100644
--- a/recipes/CommonVoice/ASR/seq2seq/hparams/train_es.yaml
+++ b/recipes/CommonVoice/ASR/seq2seq/hparams/train_es.yaml
@@ -30,7 +30,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 25
number_of_ctc_epochs: 20
lr: 1.0
@@ -61,7 +62,7 @@ sample_rate: 16000
n_fft: 400
n_mels: 80
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 3
@@ -103,50 +104,27 @@ temperature: 1.50
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
+############################## Augmentations ###################################
+ # Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 2
+ drop_count_high: 2
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -163,6 +141,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Models ##########################################
+
enc: !new:speechbrain.lobes.models.CRDNN.CRDNN
input_shape: [null, null, !ref <n_mels>]
activation: !ref <activation>
diff --git a/recipes/CommonVoice/ASR/seq2seq/hparams/train_fr.yaml b/recipes/CommonVoice/ASR/seq2seq/hparams/train_fr.yaml
index 1c2a85ec8..cc9b0aa99 100644
--- a/recipes/CommonVoice/ASR/seq2seq/hparams/train_fr.yaml
+++ b/recipes/CommonVoice/ASR/seq2seq/hparams/train_fr.yaml
@@ -29,7 +29,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 25
number_of_ctc_epochs: 20
lr: 1.0
@@ -60,7 +61,7 @@ sample_rate: 16000
n_fft: 400
n_mels: 80
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 3
@@ -102,50 +103,27 @@ temperature: 1.50
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
+############################## Augmentations ###################################
+ # Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 2
+ drop_count_high: 2
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -162,6 +140,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Models ##########################################
+
enc: !new:speechbrain.lobes.models.CRDNN.CRDNN
input_shape: [null, null, !ref <n_mels>]
activation: !ref <activation>
diff --git a/recipes/CommonVoice/ASR/seq2seq/hparams/train_it.yaml b/recipes/CommonVoice/ASR/seq2seq/hparams/train_it.yaml
index bf9211caf..2c0355ae5 100644
--- a/recipes/CommonVoice/ASR/seq2seq/hparams/train_it.yaml
+++ b/recipes/CommonVoice/ASR/seq2seq/hparams/train_it.yaml
@@ -29,7 +29,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 8.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 50
number_of_ctc_epochs: 40
lr: 1.0
@@ -59,7 +60,7 @@ sample_rate: 16000
n_fft: 400
n_mels: 80
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 3
@@ -101,50 +102,27 @@ temperature: 1.50
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
+############################## Augmentations ###################################
+ # Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 2
+ drop_count_high: 2
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -161,6 +139,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Models ##########################################
+
enc: !new:speechbrain.lobes.models.CRDNN.CRDNN
input_shape: [null, null, !ref <n_mels>]
activation: !ref <activation>
diff --git a/recipes/CommonVoice/ASR/seq2seq/hparams/train_rw.yaml b/recipes/CommonVoice/ASR/seq2seq/hparams/train_rw.yaml
index dd665ab24..8bc89c1c4 100644
--- a/recipes/CommonVoice/ASR/seq2seq/hparams/train_rw.yaml
+++ b/recipes/CommonVoice/ASR/seq2seq/hparams/train_rw.yaml
@@ -29,7 +29,8 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 8.0
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 25
number_of_ctc_epochs: 20
lr: 1.0
@@ -59,7 +60,7 @@ sample_rate: 16000
n_fft: 400
n_mels: 80
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 3
@@ -102,50 +103,27 @@ epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
+############################## Augmentations ###################################
+ # Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 2
+ drop_count_high: 2
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -162,6 +140,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Models ##########################################
+
enc: !new:speechbrain.lobes.models.CRDNN.CRDNN
input_shape: [null, null, !ref <n_mels>]
activation: !ref <activation>
diff --git a/recipes/CommonVoice/ASR/transducer/hparams/train_de.yaml b/recipes/CommonVoice/ASR/transducer/hparams/train_de.yaml
index 8faaa805f..9bbab1669 100644
--- a/recipes/CommonVoice/ASR/transducer/hparams/train_de.yaml
+++ b/recipes/CommonVoice/ASR/transducer/hparams/train_de.yaml
@@ -32,7 +32,7 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
batch_size: 6
batch_size_valid: 1
@@ -71,7 +71,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: !ref <batch_size_valid>
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 3
@@ -109,50 +109,28 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
-# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
+############################## Augmentations ###################################
+# Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 3
+ drop_count_high: 3
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -161,6 +139,8 @@ fea_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <freq_drop>,
!ref <time_warp>]
+############################## Models ##########################################
+
enc: !new:speechbrain.lobes.models.CRDNN.CRDNN
input_shape: [null, null, !ref <n_mels>]
activation: !ref <activation>
diff --git a/recipes/CommonVoice/ASR/transducer/hparams/train_fr.yaml b/recipes/CommonVoice/ASR/transducer/hparams/train_fr.yaml
index 6c3f0bc7d..c96a09394 100644
--- a/recipes/CommonVoice/ASR/transducer/hparams/train_fr.yaml
+++ b/recipes/CommonVoice/ASR/transducer/hparams/train_fr.yaml
@@ -32,7 +32,7 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
batch_size: 6
batch_size_valid: 1
@@ -71,7 +71,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: !ref <batch_size_valid>
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 3
@@ -109,50 +109,28 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
-# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
+############################## Augmentations ###################################
+# Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 3
+ drop_count_high: 3
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -161,6 +139,8 @@ fea_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <freq_drop>,
!ref <time_warp>]
+############################## Models ##########################################
+
enc: !new:speechbrain.lobes.models.CRDNN.CRDNN
input_shape: [null, null, !ref <n_mels>]
activation: !ref <activation>
diff --git a/recipes/CommonVoice/ASR/transducer/hparams/train_it.yaml b/recipes/CommonVoice/ASR/transducer/hparams/train_it.yaml
index a645f9815..cf366205e 100644
--- a/recipes/CommonVoice/ASR/transducer/hparams/train_it.yaml
+++ b/recipes/CommonVoice/ASR/transducer/hparams/train_it.yaml
@@ -32,7 +32,7 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
batch_size: 6
batch_size_valid: 1
@@ -71,7 +71,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: !ref <batch_size_valid>
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 3
@@ -109,50 +109,28 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
-# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
+############################## Augmentations ###################################
+# Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 3
+ drop_count_high: 3
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -161,6 +139,8 @@ fea_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <freq_drop>,
!ref <time_warp>]
+############################## Models ##########################################
+
enc: !new:speechbrain.lobes.models.CRDNN.CRDNN
input_shape: [null, null, !ref <n_mels>]
activation: !ref <activation>
diff --git a/recipes/CommonVoice/ASR/transducer/train.py b/recipes/CommonVoice/ASR/transducer/train.py
index 1782408b8..0304aabc8 100644
--- a/recipes/CommonVoice/ASR/transducer/train.py
+++ b/recipes/CommonVoice/ASR/transducer/train.py
@@ -134,26 +134,22 @@ class ASR(sb.Brain):
if stage == sb.Stage.TRAIN:
if hasattr(self.hparams, "wav_augment"):
- tokens = self.hparams.wav_augment.replicate_labels(tokens)
- token_lens = self.hparams.wav_augment.replicate_labels(
- token_lens
- )
- tokens_eos = self.hparams.wav_augment.replicate_labels(
- tokens_eos
- )
- token_eos_lens = self.hparams.wav_augment.replicate_labels(
- token_eos_lens
+ (
+ tokens,
+ token_lens,
+ tokens_eos,
+ token_eos_lens,
+ ) = self.hparams.wav_augment.replicate_multiple_labels(
+ tokens, token_lens, tokens_eos, token_eos_lens
)
if hasattr(self.hparams, "fea_augment"):
- tokens = self.hparams.fea_augment.replicate_labels(tokens)
- token_lens = self.hparams.fea_augment.replicate_labels(
- token_lens
- )
- tokens_eos = self.hparams.fea_augment.replicate_labels(
- tokens_eos
- )
- token_eos_lens = self.hparams.fea_augment.replicate_labels(
- token_eos_lens
+ (
+ tokens,
+ token_lens,
+ tokens_eos,
+ token_eos_lens,
+ ) = self.hparams.fea_augment.replicate_multiple_labels(
+ tokens, token_lens, tokens_eos, token_eos_lens
)
if stage == sb.Stage.TRAIN:
diff --git a/recipes/CommonVoice/ASR/transformer/hparams/train_ar_hf_whisper.yaml b/recipes/CommonVoice/ASR/transformer/hparams/train_ar_hf_whisper.yaml
index 2b358e6d4..d33c50c2b 100644
--- a/recipes/CommonVoice/ASR/transformer/hparams/train_ar_hf_whisper.yaml
+++ b/recipes/CommonVoice/ASR/transformer/hparams/train_ar_hf_whisper.yaml
@@ -37,7 +37,7 @@ avoid_if_longer_than: 10.0
ckpt_interval_minutes: 30 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 1
lr_whisper: 0.00003
sorting: ascending
@@ -63,7 +63,7 @@ min_decode_ratio: 0.0
max_decode_ratio: 1.0
test_beam_size: 8
-# Model parameters
+####################### Model Parameters #######################################
freeze_whisper: False
freeze_encoder: True
@@ -82,45 +82,30 @@ test_loader_kwargs:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -129,6 +114,7 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
whisper: !new:speechbrain.lobes.models.huggingface_transformers.whisper.Whisper
source: !ref <whisper_hub>
diff --git a/recipes/CommonVoice/ASR/transformer/hparams/train_de.yaml b/recipes/CommonVoice/ASR/transformer/hparams/train_de.yaml
index dbeb56be5..c5533e9bb 100644
--- a/recipes/CommonVoice/ASR/transformer/hparams/train_de.yaml
+++ b/recipes/CommonVoice/ASR/transformer/hparams/train_de.yaml
@@ -33,7 +33,7 @@ avoid_if_longer_than: 10.0
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 50
batch_size: 32 # This works with a 32GB GPU ! (bs * nb_gpu * accum) > 128 !
ctc_weight: 0.3
@@ -70,7 +70,7 @@ test_dataloader_opts:
batch_size: !ref <batch_size>
num_workers: 6
-####################### Model parameters ###########################
+####################### Model Parameters ###########################
# Transformer
d_model: 768
nhead: 8
@@ -213,50 +213,27 @@ normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
update_until_epoch: 3
-# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
+############################## Augmentations ###################################
+# Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 2
+ drop_count_high: 2
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
diff --git a/recipes/CommonVoice/ASR/transformer/hparams/train_fa_hf_whisper.yaml b/recipes/CommonVoice/ASR/transformer/hparams/train_fa_hf_whisper.yaml
index bc0181044..bb23c98a6 100644
--- a/recipes/CommonVoice/ASR/transformer/hparams/train_fa_hf_whisper.yaml
+++ b/recipes/CommonVoice/ASR/transformer/hparams/train_fa_hf_whisper.yaml
@@ -37,7 +37,7 @@ avoid_if_longer_than: 10.0
ckpt_interval_minutes: 30 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 1
lr_whisper: 0.00003
sorting: ascending
@@ -63,7 +63,7 @@ min_decode_ratio: 0.0
max_decode_ratio: 1.0
test_beam_size: 8
-# Model parameters
+####################### Model Parameters #######################################
freeze_whisper: False
freeze_encoder: True
@@ -82,45 +82,30 @@ test_loader_kwargs:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -129,6 +114,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
whisper: !new:speechbrain.lobes.models.huggingface_transformers.whisper.Whisper
source: !ref <whisper_hub>
diff --git a/recipes/CommonVoice/ASR/transformer/hparams/train_fr.yaml b/recipes/CommonVoice/ASR/transformer/hparams/train_fr.yaml
index 120305e73..e62d9c390 100644
--- a/recipes/CommonVoice/ASR/transformer/hparams/train_fr.yaml
+++ b/recipes/CommonVoice/ASR/transformer/hparams/train_fr.yaml
@@ -33,7 +33,7 @@ avoid_if_longer_than: 10.0
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 50
batch_size: 32 # This works with a 32GB GPU ! (bs * nb_gpu * accum) > 128 !
ctc_weight: 0.3
@@ -70,7 +70,7 @@ test_dataloader_opts:
batch_size: !ref <batch_size>
num_workers: 6
-####################### Model parameters ###########################
+####################### Model Parameters ###########################
# Transformer
d_model: 768
nhead: 8
@@ -213,50 +213,27 @@ normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
update_until_epoch: 3
-# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
+############################## Augmentations ###################################
+# Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 2
+ drop_count_high: 2
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
diff --git a/recipes/CommonVoice/ASR/transformer/hparams/train_fr_hf_whisper.yaml b/recipes/CommonVoice/ASR/transformer/hparams/train_fr_hf_whisper.yaml
index da5cbd28f..62363bdad 100644
--- a/recipes/CommonVoice/ASR/transformer/hparams/train_fr_hf_whisper.yaml
+++ b/recipes/CommonVoice/ASR/transformer/hparams/train_fr_hf_whisper.yaml
@@ -37,7 +37,7 @@ avoid_if_longer_than: 10.0
ckpt_interval_minutes: 30 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 1
lr_whisper: 0.00003
sorting: ascending
@@ -63,7 +63,7 @@ min_decode_ratio: 0.0
max_decode_ratio: 1.0
test_beam_size: 8
-# Model parameters
+####################### Model Parameters #######################################
freeze_whisper: False
freeze_encoder: True
@@ -82,45 +82,30 @@ test_loader_kwargs:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -129,6 +114,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
whisper: !new:speechbrain.lobes.models.huggingface_transformers.whisper.Whisper
source: !ref <whisper_hub>
diff --git a/recipes/CommonVoice/ASR/transformer/hparams/train_hi_hf_whisper.yaml b/recipes/CommonVoice/ASR/transformer/hparams/train_hi_hf_whisper.yaml
index 8b130a83d..e21852639 100644
--- a/recipes/CommonVoice/ASR/transformer/hparams/train_hi_hf_whisper.yaml
+++ b/recipes/CommonVoice/ASR/transformer/hparams/train_hi_hf_whisper.yaml
@@ -37,7 +37,7 @@ avoid_if_longer_than: 10.0
ckpt_interval_minutes: 30 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 1
lr_whisper: 0.00003
sorting: ascending
@@ -63,7 +63,7 @@ min_decode_ratio: 0.0
max_decode_ratio: 1.0
test_beam_size: 8
-# Model parameters
+####################### Model Parameters #######################################
freeze_whisper: False
freeze_encoder: True
@@ -82,45 +82,30 @@ test_loader_kwargs:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -129,6 +114,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
whisper: !new:speechbrain.lobes.models.huggingface_transformers.whisper.Whisper
source: !ref <whisper_hub>
diff --git a/recipes/CommonVoice/ASR/transformer/hparams/train_it.yaml b/recipes/CommonVoice/ASR/transformer/hparams/train_it.yaml
index d937359b5..d95fbaffa 100644
--- a/recipes/CommonVoice/ASR/transformer/hparams/train_it.yaml
+++ b/recipes/CommonVoice/ASR/transformer/hparams/train_it.yaml
@@ -33,7 +33,7 @@ avoid_if_longer_than: 10.0
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 50
batch_size: 32 # This works with a 32GB GPU ! (bs * nb_gpu * accum) > 128 !
ctc_weight: 0.3
@@ -70,7 +70,7 @@ test_dataloader_opts:
batch_size: !ref <batch_size>
num_workers: 6
-####################### Model parameters ###########################
+####################### Model Parameters ###########################
# Transformer
d_model: 768
nhead: 8
@@ -213,50 +213,27 @@ normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
update_until_epoch: 3
-# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
+############################## Augmentations ###################################
+# Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 2
+ drop_count_high: 2
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
diff --git a/recipes/CommonVoice/ASR/transformer/hparams/train_it_hf_whisper.yaml b/recipes/CommonVoice/ASR/transformer/hparams/train_it_hf_whisper.yaml
index 5670f4fe8..e1fc08263 100644
--- a/recipes/CommonVoice/ASR/transformer/hparams/train_it_hf_whisper.yaml
+++ b/recipes/CommonVoice/ASR/transformer/hparams/train_it_hf_whisper.yaml
@@ -37,7 +37,7 @@ avoid_if_longer_than: 10.0
ckpt_interval_minutes: 30 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 1
lr_whisper: 0.00003
sorting: ascending
@@ -63,7 +63,7 @@ min_decode_ratio: 0.0
max_decode_ratio: 1.0
test_beam_size: 8
-# Model parameters
+####################### Model Parameters #######################################
freeze_whisper: False
freeze_encoder: True
@@ -82,45 +82,30 @@ test_loader_kwargs:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -129,6 +114,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
whisper: !new:speechbrain.lobes.models.huggingface_transformers.whisper.Whisper
source: !ref <whisper_hub>
diff --git a/recipes/CommonVoice/ASR/transformer/hparams/train_mn_hf_whisper.yaml b/recipes/CommonVoice/ASR/transformer/hparams/train_mn_hf_whisper.yaml
index 9ffdc95fd..fe4fd6f17 100644
--- a/recipes/CommonVoice/ASR/transformer/hparams/train_mn_hf_whisper.yaml
+++ b/recipes/CommonVoice/ASR/transformer/hparams/train_mn_hf_whisper.yaml
@@ -37,7 +37,7 @@ avoid_if_longer_than: 10.0
ckpt_interval_minutes: 30 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 1
lr_whisper: 0.00003
sorting: ascending
@@ -63,7 +63,7 @@ min_decode_ratio: 0.0
max_decode_ratio: 1.0
test_beam_size: 8
-# Model parameters
+####################### Model Parameters #######################################
freeze_whisper: False
freeze_encoder: True
@@ -82,45 +82,30 @@ test_loader_kwargs:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -129,6 +114,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
whisper: !new:speechbrain.lobes.models.huggingface_transformers.whisper.Whisper
source: !ref <whisper_hub>
diff --git a/recipes/CommonVoice/ASR/transformer/hparams/train_sr_hf_whisper.yaml b/recipes/CommonVoice/ASR/transformer/hparams/train_sr_hf_whisper.yaml
index 4f257094d..d7390d9a5 100644
--- a/recipes/CommonVoice/ASR/transformer/hparams/train_sr_hf_whisper.yaml
+++ b/recipes/CommonVoice/ASR/transformer/hparams/train_sr_hf_whisper.yaml
@@ -37,7 +37,7 @@ avoid_if_longer_than: 10.0
ckpt_interval_minutes: 30 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 1
lr_whisper: 0.00003
sorting: ascending
@@ -63,7 +63,7 @@ min_decode_ratio: 0.0
max_decode_ratio: 1.0
test_beam_size: 8
-# Model parameters
+####################### Model Parameters #######################################
freeze_whisper: False
freeze_encoder: True
@@ -83,45 +83,30 @@ test_loader_kwargs:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -130,6 +115,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
whisper: !new:speechbrain.lobes.models.huggingface_transformers.whisper.Whisper
source: !ref <whisper_hub>
diff --git a/recipes/CommonVoice/ASR/transformer/train.py b/recipes/CommonVoice/ASR/transformer/train.py
index 0aee6735e..89847d352 100644
--- a/recipes/CommonVoice/ASR/transformer/train.py
+++ b/recipes/CommonVoice/ASR/transformer/train.py
@@ -107,27 +107,25 @@ class ASR(sb.core.Brain):
# Augment Labels
if stage == sb.Stage.TRAIN:
+ # Labels must be extended if parallel augmentation or concatenated
+ # augmentation was performed on the input (increasing the time dimension)
if hasattr(self.hparams, "wav_augment"):
- tokens = self.hparams.wav_augment.replicate_labels(tokens)
- tokens_lens = self.hparams.wav_augment.replicate_labels(
- tokens_lens
- )
- tokens_eos = self.hparams.wav_augment.replicate_labels(
- tokens_eos
- )
- tokens_eos_lens = self.hparams.wav_augment.replicate_labels(
- tokens_eos_lens
+ (
+ tokens,
+ tokens_lens,
+ tokens_eos,
+ tokens_eos_lens,
+ ) = self.hparams.wav_augment.replicate_multiple_labels(
+ tokens, tokens_lens, tokens_eos, tokens_eos_lens
)
if hasattr(self.hparams, "fea_augment"):
- tokens = self.hparams.fea_augment.replicate_labels(tokens)
- tokens_lens = self.hparams.fea_augment.replicate_labels(
- tokens_lens
- )
- tokens_eos = self.hparams.fea_augment.replicate_labels(
- tokens_eos
- )
- tokens_eos_lens = self.hparams.fea_augment.replicate_labels(
- tokens_eos_lens
+ (
+ tokens,
+ tokens_lens,
+ tokens_eos,
+ tokens_eos_lens,
+ ) = self.hparams.fea_augment.replicate_multiple_labels(
+ tokens, tokens_lens, tokens_eos, tokens_eos_lens
)
loss_seq = self.hparams.seq_cost(
diff --git a/recipes/CommonVoice/self-supervised-learning/wav2vec2/hparams/wav2vec2_base.yaml b/recipes/CommonVoice/self-supervised-learning/wav2vec2/hparams/wav2vec2_base.yaml
index 5ead6dbc0..e7ceed4f5 100644
--- a/recipes/CommonVoice/self-supervised-learning/wav2vec2/hparams/wav2vec2_base.yaml
+++ b/recipes/CommonVoice/self-supervised-learning/wav2vec2/hparams/wav2vec2_base.yaml
@@ -27,11 +27,11 @@ skip_prep: False
# We remove utterance slonger than 10s in the train/dev/test sets as
-# longer sentences certainly correspond to "open microphones".
+# longer sentences certainly correspond to open microphones.
avoid_if_longer_than: 10.0
avoid_if_shorter_than: 1.0
-# Training parameters
+####################### Training Parameters ####################################
# Parameters are corresponding the the ones reported in the official wav2vec2
# paper (for the masking).
mask_length: 10
@@ -52,8 +52,8 @@ ckpt_interval_minutes: 30 # save checkpoint every N min
# IMPORTANT: To train w2v2 model, we recommand to have the effective batch_size
# higher than 100 (batch_size * nb_gpu * grad_accumulation_factor)
# Examples are:
-# 32 Tesla V100 32GB — 12 * 32 * 1
-# 4 Tesla V100 32GB — 12 * 4 * {6-8}
+# 32 Tesla V100 32GB = 12 * 32 * 1
+# 4 Tesla V100 32GB = 12 * 4 * (6-8)
batch_size: 12
test_batch_size: 8
grad_accumulation_factor: 8
@@ -104,7 +104,7 @@ modules:
wav2vec2: !ref <wav2vec2>
opt_class: !name:torch.optim.AdamW
- lr: 0 # Will be changed by the scheduler, but we start at 0!
+ lr: 0 # Will be changed by the scheduler, but we start at 0
betas: (0.9, 0.98)
eps: 0.000000001
weight_decay: !ref <weight_decay>
diff --git a/recipes/DNS/enhancement/hparams/sepformer-dns-16k.yaml b/recipes/DNS/enhancement/hparams/sepformer-dns-16k.yaml
index d43a2f0a9..87a07c97a 100644
--- a/recipes/DNS/enhancement/hparams/sepformer-dns-16k.yaml
+++ b/recipes/DNS/enhancement/hparams/sepformer-dns-16k.yaml
@@ -39,7 +39,7 @@ sample_rate: 16000
audio_length: 4 # seconds
n_audio_to_save: 20
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 100
batch_size: 4
batch_size_test: 1
diff --git a/recipes/DVoice/ASR/CTC/hparams/train_amh_with_wav2vec.yaml b/recipes/DVoice/ASR/CTC/hparams/train_amh_with_wav2vec.yaml
index 551ba2c19..e9e1f4310 100644
--- a/recipes/DVoice/ASR/CTC/hparams/train_amh_with_wav2vec.yaml
+++ b/recipes/DVoice/ASR/CTC/hparams/train_amh_with_wav2vec.yaml
@@ -32,7 +32,7 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 15.0
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -58,7 +58,7 @@ test_dataloader_options:
token_type: char # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
wav2vec_output_dim: 1024
dnn_neurons: 1024
freeze_wav2vec: False
@@ -78,45 +78,31 @@ eos_index: 2
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -125,6 +111,7 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
diff --git a/recipes/DVoice/ASR/CTC/hparams/train_dar_with_wav2vec.yaml b/recipes/DVoice/ASR/CTC/hparams/train_dar_with_wav2vec.yaml
index 0b3647705..d1e2c6684 100644
--- a/recipes/DVoice/ASR/CTC/hparams/train_dar_with_wav2vec.yaml
+++ b/recipes/DVoice/ASR/CTC/hparams/train_dar_with_wav2vec.yaml
@@ -32,7 +32,7 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 15.0
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -58,7 +58,7 @@ test_dataloader_options:
token_type: char # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
wav2vec_output_dim: 1024
dnn_neurons: 1024
freeze_wav2vec: False
@@ -78,45 +78,31 @@ eos_index: 2
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -125,6 +111,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
diff --git a/recipes/DVoice/ASR/CTC/hparams/train_fon_with_wav2vec.yaml b/recipes/DVoice/ASR/CTC/hparams/train_fon_with_wav2vec.yaml
index 946ca0b6f..fca0230de 100644
--- a/recipes/DVoice/ASR/CTC/hparams/train_fon_with_wav2vec.yaml
+++ b/recipes/DVoice/ASR/CTC/hparams/train_fon_with_wav2vec.yaml
@@ -32,7 +32,7 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 15.0
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -58,7 +58,7 @@ test_dataloader_options:
token_type: char # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
wav2vec_output_dim: 1024
dnn_neurons: 1024
freeze_wav2vec: False
@@ -78,45 +78,31 @@ eos_index: 2
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -125,6 +111,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
diff --git a/recipes/DVoice/ASR/CTC/hparams/train_multi_with_wav2vec.yaml b/recipes/DVoice/ASR/CTC/hparams/train_multi_with_wav2vec.yaml
index 14aef36c9..89fedade8 100644
--- a/recipes/DVoice/ASR/CTC/hparams/train_multi_with_wav2vec.yaml
+++ b/recipes/DVoice/ASR/CTC/hparams/train_multi_with_wav2vec.yaml
@@ -31,7 +31,7 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 15.0
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -57,7 +57,7 @@ test_dataloader_options:
token_type: char # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
wav2vec_output_dim: 1024
dnn_neurons: 1024
freeze_wav2vec: False
@@ -77,45 +77,31 @@ eos_index: 2
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -124,6 +110,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
diff --git a/recipes/DVoice/ASR/CTC/hparams/train_sw_with_wav2vec.yaml b/recipes/DVoice/ASR/CTC/hparams/train_sw_with_wav2vec.yaml
index f00e330a4..0194fd877 100644
--- a/recipes/DVoice/ASR/CTC/hparams/train_sw_with_wav2vec.yaml
+++ b/recipes/DVoice/ASR/CTC/hparams/train_sw_with_wav2vec.yaml
@@ -32,7 +32,7 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 15.0
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -58,7 +58,7 @@ test_dataloader_options:
token_type: char # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
wav2vec_output_dim: 1024
dnn_neurons: 1024
freeze_wav2vec: False
@@ -78,45 +78,31 @@ eos_index: 2
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -125,6 +111,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
diff --git a/recipes/DVoice/ASR/CTC/hparams/train_wol_with_wav2vec.yaml b/recipes/DVoice/ASR/CTC/hparams/train_wol_with_wav2vec.yaml
index b1188dcb9..8470ce3a1 100644
--- a/recipes/DVoice/ASR/CTC/hparams/train_wol_with_wav2vec.yaml
+++ b/recipes/DVoice/ASR/CTC/hparams/train_wol_with_wav2vec.yaml
@@ -32,7 +32,7 @@ skip_prep: False # Skip data preparation
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 15.0
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -58,7 +58,7 @@ test_dataloader_options:
token_type: char # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
wav2vec_output_dim: 1024
dnn_neurons: 1024
freeze_wav2vec: False
@@ -78,45 +78,31 @@ eos_index: 2
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -125,6 +111,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
linear1: !name:speechbrain.nnet.linear.Linear
diff --git a/recipes/ESC50/classification/hparams/cnn14_classifier.yaml b/recipes/ESC50/classification/hparams/cnn14_classifier.yaml
index e8034bfdd..bc0a83bbd 100644
--- a/recipes/ESC50/classification/hparams/cnn14_classifier.yaml
+++ b/recipes/ESC50/classification/hparams/cnn14_classifier.yaml
@@ -41,7 +41,7 @@ skip_manifest_creation: False
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 200
batch_size: 32
lr: 0.0002
diff --git a/recipes/ESC50/classification/hparams/conv2d_classifier.yaml b/recipes/ESC50/classification/hparams/conv2d_classifier.yaml
index 2b0a49bcd..284d5681f 100644
--- a/recipes/ESC50/classification/hparams/conv2d_classifier.yaml
+++ b/recipes/ESC50/classification/hparams/conv2d_classifier.yaml
@@ -41,7 +41,7 @@ skip_manifest_creation: False
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 200
batch_size: 32
lr: 0.00002
diff --git a/recipes/ESC50/interpret/hparams/l2i_cnn14.yaml b/recipes/ESC50/interpret/hparams/l2i_cnn14.yaml
index 6f57a843b..00acd1ff3 100644
--- a/recipes/ESC50/interpret/hparams/l2i_cnn14.yaml
+++ b/recipes/ESC50/interpret/hparams/l2i_cnn14.yaml
@@ -39,7 +39,7 @@ skip_manifest_creation: False
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 200
batch_size: 2
lr: 0.0001
diff --git a/recipes/ESC50/interpret/hparams/l2i_conv2dclassifier.yaml b/recipes/ESC50/interpret/hparams/l2i_conv2dclassifier.yaml
index 7292f89af..4f6cb9b90 100644
--- a/recipes/ESC50/interpret/hparams/l2i_conv2dclassifier.yaml
+++ b/recipes/ESC50/interpret/hparams/l2i_conv2dclassifier.yaml
@@ -39,7 +39,7 @@ skip_manifest_creation: False
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 200
batch_size: 16
lr: 0.0002
diff --git a/recipes/ESC50/interpret/hparams/nmf.yaml b/recipes/ESC50/interpret/hparams/nmf.yaml
index 7b6c9905d..e4da313ba 100644
--- a/recipes/ESC50/interpret/hparams/nmf.yaml
+++ b/recipes/ESC50/interpret/hparams/nmf.yaml
@@ -40,7 +40,7 @@ skip_manifest_creation: False
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 100
batch_size: 2
lr: 0.0002
diff --git a/recipes/ESC50/interpret/hparams/piq.yaml b/recipes/ESC50/interpret/hparams/piq.yaml
index c45f50a20..68f8c06de 100644
--- a/recipes/ESC50/interpret/hparams/piq.yaml
+++ b/recipes/ESC50/interpret/hparams/piq.yaml
@@ -42,7 +42,7 @@ skip_manifest_creation: False
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 200
batch_size: 16
lr: 0.0002
diff --git a/recipes/Fisher-Callhome-Spanish/ST/transformer/hparams/conformer.yaml b/recipes/Fisher-Callhome-Spanish/ST/transformer/hparams/conformer.yaml
index ec8653ade..49a7321f7 100644
--- a/recipes/Fisher-Callhome-Spanish/ST/transformer/hparams/conformer.yaml
+++ b/recipes/Fisher-Callhome-Spanish/ST/transformer/hparams/conformer.yaml
@@ -81,7 +81,7 @@ test_dataloader_opts:
batch_size: !ref <batch_size>
num_workers: !ref <num_workers>
-####################### Model parameters ###########################
+####################### Model Parameters ###########################
# Transformer
d_model: 256
nhead: 4
diff --git a/recipes/Fisher-Callhome-Spanish/ST/transformer/hparams/transformer.yaml b/recipes/Fisher-Callhome-Spanish/ST/transformer/hparams/transformer.yaml
index 59c3782e1..4310e2d6b 100644
--- a/recipes/Fisher-Callhome-Spanish/ST/transformer/hparams/transformer.yaml
+++ b/recipes/Fisher-Callhome-Spanish/ST/transformer/hparams/transformer.yaml
@@ -91,7 +91,7 @@ test_dataloader_opts:
batch_size: !ref <batch_size>
num_workers: !ref <num_workers>
-####################### Model parameters ###########################
+####################### Model Parameters ###########################
# Transformer
d_model: 256
nhead: 4
diff --git a/recipes/Google-speech-commands/hparams/xvect.yaml b/recipes/Google-speech-commands/hparams/xvect.yaml
index 8eb842ba9..417cecfdf 100644
--- a/recipes/Google-speech-commands/hparams/xvect.yaml
+++ b/recipes/Google-speech-commands/hparams/xvect.yaml
@@ -40,7 +40,7 @@ percentage_silence: 10 # Set this to 0 for the V2 35 task
skip_prep: False
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 100
batch_size: 32
lr: 0.001
diff --git a/recipes/Google-speech-commands/hparams/xvect_leaf.yaml b/recipes/Google-speech-commands/hparams/xvect_leaf.yaml
index e06101850..f2897af22 100644
--- a/recipes/Google-speech-commands/hparams/xvect_leaf.yaml
+++ b/recipes/Google-speech-commands/hparams/xvect_leaf.yaml
@@ -42,7 +42,7 @@ percentage_silence: 10 # Set this to 0 for the V2 35 task
skip_prep: False
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 100
batch_size: 32
lr: 0.001
diff --git a/recipes/IEMOCAP/emotion_recognition/hparams/train_with_wav2vec2.yaml b/recipes/IEMOCAP/emotion_recognition/hparams/train_with_wav2vec2.yaml
index ae3452882..d1b63d7bf 100644
--- a/recipes/IEMOCAP/emotion_recognition/hparams/train_with_wav2vec2.yaml
+++ b/recipes/IEMOCAP/emotion_recognition/hparams/train_with_wav2vec2.yaml
@@ -38,7 +38,7 @@ train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
batch_size: 4
lr: 0.0001
@@ -50,7 +50,7 @@ freeze_wav2vec2: False
# We see an improvement of 2% with freezing CNNs
freeze_wav2vec2_conv: True
-# Model parameters
+####################### Model Parameters #######################################
encoder_dim: 768
# Number of emotions
diff --git a/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_samu.yaml b/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_samu.yaml
index 05bca5c1e..3901391a5 100644
--- a/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_samu.yaml
+++ b/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_samu.yaml
@@ -30,7 +30,7 @@ wav2vec2_hub: LIA-AvignonUniversity/IWSLT2022-tamasheq-only
# wav2vec 2.0 specific parameters
wav2vec2_frozen: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 100
lr: 0.001
lr_wav2vec: 0.00001
diff --git a/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_samu_mbart_st.yaml b/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_samu_mbart_st.yaml
index a3a2f1c99..6887c3a40 100644
--- a/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_samu_mbart_st.yaml
+++ b/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_samu_mbart_st.yaml
@@ -36,7 +36,7 @@ wav2vec2_folder: !ref <save_folder>/wav2vec2_checkpoint
# wav2vec 2.0 specific parameters
wav2vec2_frozen: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 500
lr: 0.001
lr_wav2vec: 0.0001
diff --git a/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_samu_nllb_st.yaml b/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_samu_nllb_st.yaml
index 11ebc937a..b86cef685 100644
--- a/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_samu_nllb_st.yaml
+++ b/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_samu_nllb_st.yaml
@@ -36,7 +36,7 @@ wav2vec2_folder: !ref <save_folder>/wav2vec2_checkpoint
# wav2vec 2.0 specific parameters
wav2vec2_frozen: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 500
lr: 0.001
lr_wav2vec: 0.0001
diff --git a/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_w2v2_mbart_st.yaml b/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_w2v2_mbart_st.yaml
index 68f74d9b4..77b7c8cd6 100644
--- a/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_w2v2_mbart_st.yaml
+++ b/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_w2v2_mbart_st.yaml
@@ -36,7 +36,7 @@ wav2vec2_folder: !ref <save_folder>/wav2vec2_checkpoint
# wav2vec 2.0 specific parameters
wav2vec2_frozen: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 500
lr: 0.001
lr_wav2vec: 0.0001
diff --git a/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_w2v2_nllb_st.yaml b/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_w2v2_nllb_st.yaml
index b62e366f1..d384bf3a8 100644
--- a/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_w2v2_nllb_st.yaml
+++ b/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_w2v2_nllb_st.yaml
@@ -36,7 +36,7 @@ wav2vec2_folder: !ref <save_folder>/wav2vec2_checkpoint
# wav2vec 2.0 specific parameters
wav2vec2_frozen: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 500
lr: 0.001
lr_wav2vec: 0.0001
diff --git a/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_w2v2_st.yaml b/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_w2v2_st.yaml
index 6bfb9db12..beafeba86 100644
--- a/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_w2v2_st.yaml
+++ b/recipes/IWSLT22_lowresource/AST/transformer/hparams/train_w2v2_st.yaml
@@ -36,7 +36,7 @@ wav2vec2_folder: !ref <save_folder>/wav2vec2_checkpoint
wav2vec2_frozen: False
keep_n_layers: 6 # keep first N layers from the Transformer Encoder stack inside the wav2vec 2.0 model
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 100
lr: 0.001
lr_wav2vec: 0.00001
diff --git a/recipes/KsponSpeech/ASR/transformer/hparams/conformer_medium.yaml b/recipes/KsponSpeech/ASR/transformer/hparams/conformer_medium.yaml
index aee256973..3c0d43e2a 100644
--- a/recipes/KsponSpeech/ASR/transformer/hparams/conformer_medium.yaml
+++ b/recipes/KsponSpeech/ASR/transformer/hparams/conformer_medium.yaml
@@ -34,7 +34,7 @@ test_csv:
ckpt_interval_minutes: 30 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
# To make Transformers converge, the global bath size should be large enough.
# The global batch size is computed as batch_size * n_gpus * grad_accumulation_factor.
# Empirically, we found that this value should be >= 128.
@@ -78,7 +78,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-####################### Model parameters ###########################
+####################### Model Parameters ###########################
# Transformer
d_model: 256
nhead: 4
diff --git a/recipes/KsponSpeech/LM/hparams/transformer.yaml b/recipes/KsponSpeech/LM/hparams/transformer.yaml
index cd9685e28..5b64cc196 100644
--- a/recipes/KsponSpeech/LM/hparams/transformer.yaml
+++ b/recipes/KsponSpeech/LM/hparams/transformer.yaml
@@ -24,7 +24,7 @@ test_csv:
# Tokenizer model
tokenizer_file: ddwkim/asr-conformer-transformerlm-ksponspeech/tokenizer.ckpt
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
batch_size: 256
lr: 0.1
diff --git a/recipes/KsponSpeech/Tokenizer/hparams/5K_unigram_subword_bpe.yaml b/recipes/KsponSpeech/Tokenizer/hparams/5K_unigram_subword_bpe.yaml
index dd7cd4906..04ef0ebfd 100644
--- a/recipes/KsponSpeech/Tokenizer/hparams/5K_unigram_subword_bpe.yaml
+++ b/recipes/KsponSpeech/Tokenizer/hparams/5K_unigram_subword_bpe.yaml
@@ -16,7 +16,7 @@ skip_prep: False
train_csv: !ref <output_folder>/train.csv
valid_csv: !ref <output_folder>/dev.csv
-# Training parameters
+####################### Training Parameters ####################################
token_type: unigram # ["unigram", "bpe", "char"]
token_output: 5000 # index(blank/eos/bos/unk) = 0
character_coverage: 1.0
diff --git a/recipes/LibriMix/separation/hparams/sepformer-libri2mix.yaml b/recipes/LibriMix/separation/hparams/sepformer-libri2mix.yaml
index 8fb195c03..ffa5a1ef2 100644
--- a/recipes/LibriMix/separation/hparams/sepformer-libri2mix.yaml
+++ b/recipes/LibriMix/separation/hparams/sepformer-libri2mix.yaml
@@ -37,7 +37,7 @@ noprogressbar: False
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/LibriMix/separation/hparams/sepformer-libri3mix.yaml b/recipes/LibriMix/separation/hparams/sepformer-libri3mix.yaml
index cf68e9a81..abc9c76c7 100644
--- a/recipes/LibriMix/separation/hparams/sepformer-libri3mix.yaml
+++ b/recipes/LibriMix/separation/hparams/sepformer-libri3mix.yaml
@@ -37,7 +37,7 @@ noprogressbar: False
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/LibriParty/VAD/hparams/train.yaml b/recipes/LibriParty/VAD/hparams/train.yaml
index e07258c44..be9191685 100644
--- a/recipes/LibriParty/VAD/hparams/train.yaml
+++ b/recipes/LibriParty/VAD/hparams/train.yaml
@@ -41,7 +41,7 @@ speech_csv: !ref <save_folder>/speech.csv
multilang_speech_csv: !ref <save_folder>/multilang_speech.csv
skip_prep: False # Skip data preparation
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 100
lr: 1.0
lr_final: 0.1
@@ -65,7 +65,7 @@ test_dataloader_opts:
n_fft: 400
n_mels: 40
-# Model parameters
+####################### Model Parameters #######################################
# activation: !name:torch.nn.LeakyReLU
# dropout: 0.15
# cnn_blocks: 2
diff --git a/recipes/LibriSpeech/ASR/CTC/hparams/downsampled/train_hf_wavlm_average_downsampling.yaml b/recipes/LibriSpeech/ASR/CTC/hparams/downsampled/train_hf_wavlm_average_downsampling.yaml
index b609bb766..fdbd7e86d 100644
--- a/recipes/LibriSpeech/ASR/CTC/hparams/downsampled/train_hf_wavlm_average_downsampling.yaml
+++ b/recipes/LibriSpeech/ASR/CTC/hparams/downsampled/train_hf_wavlm_average_downsampling.yaml
@@ -33,7 +33,8 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 1
lr: 0.9
lr_wav2vec: 0.0001
@@ -62,7 +63,8 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: !ref <test_batch_size>
-# Model parameters
+####################### Model Parameters #######################################
+
activation: !name:torch.nn.LeakyReLU
dnn_layers: 2
dnn_neurons: 1024
@@ -73,76 +75,12 @@ ctc_neurons: 29
output_neurons: 29 # Characters size, index(blank/eos/bos) = 0
blank_index: 0
-# Decoding parameters
-test_beam_search:
- beam_size: 200
- topk: 1
- blank_index: !ref <blank_index>
- space_token: ' ' # make sure this is the same as the one used in the tokenizer
- beam_prune_logp: -10.0
- token_prune_min_logp: -5
- prune_history: True
- alpha: 0.5
- beta: 1.5
- # can be downloaded from here https://www.openslr.org/11/ or trained with kenLM
- # It can either be a .bin or .arpa ; note: .arpa is much slower at loading
- # If you don't want to use an LM, comment it out or set it to null
- kenlm_model_path: null
-
#
# Functions and classes
#
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
-speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
- orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
-
-# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
-drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
-
-# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
-drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
-
-# Augmenter: Combines previously defined augmentations to perform data augmentation
-wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
- min_augmentations: 4
- max_augmentations: 4
- augment_prob: 1.0
- augmentations: [
- !ref <speed_perturb>,
- !ref <drop_freq>,
- !ref <drop_chunk>]
-
-
enc: !new:speechbrain.lobes.models.VanillaNN.VanillaNN
input_shape: [null, null, 1024]
activation: !ref <activation>
@@ -211,8 +149,60 @@ lr_annealing_wav2vec: !new:speechbrain.nnet.schedulers.NewBobScheduler
annealing_factor: 0.9
patient: 0
+############################## Decoding ########################################
+
+test_beam_search:
+ beam_size: 200
+ topk: 1
+ blank_index: !ref <blank_index>
+ space_token: ' ' # make sure this is the same as the one used in the tokenizer
+ beam_prune_logp: -10.0
+ token_prune_min_logp: -5
+ prune_history: True
+ alpha: 0.5
+ beta: 1.5
+ # can be downloaded from here https://www.openslr.org/11/ or trained with kenLM
+ # It can either be a .bin or .arpa ; note: .arpa is much slower at loading
+ # If you don't want to use an LM, comment it out or set it to null
+ kenlm_model_path: null
+
label_encoder: !new:speechbrain.dataio.encoder.CTCTextEncoder
+############################## Augmentations ###################################
+
+# Speed perturbation
+speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
+ orig_freq: !ref <sample_rate>
+ speeds: [95, 100, 105]
+
+# Frequency drop: randomly drops a number of frequency bands to zero.
+drop_freq: !new:speechbrain.augment.time_domain.DropFreq
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
+
+# Time drop: randomly drops a number of temporal chunks.
+drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
+
+# Augmenter: Combines previously defined augmentations to perform data augmentation
+wav_augment: !new:speechbrain.augment.augmenter.Augmenter
+ concat_original: True
+ min_augmentations: 4
+ max_augmentations: 4
+ augment_prob: 1.0
+ augmentations: [
+ !ref <speed_perturb>,
+ !ref <drop_freq>,
+ !ref <drop_chunk>]
+
+############################## Logging and Pretrainer ##########################
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
recoverables:
diff --git a/recipes/LibriSpeech/ASR/CTC/hparams/downsampled/train_hf_wavlm_conv_downsampling.yaml b/recipes/LibriSpeech/ASR/CTC/hparams/downsampled/train_hf_wavlm_conv_downsampling.yaml
index f92f4f8fc..1b84596dc 100644
--- a/recipes/LibriSpeech/ASR/CTC/hparams/downsampled/train_hf_wavlm_conv_downsampling.yaml
+++ b/recipes/LibriSpeech/ASR/CTC/hparams/downsampled/train_hf_wavlm_conv_downsampling.yaml
@@ -34,7 +34,8 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 1
lr: 0.9
lr_wav2vec: 0.0001
@@ -63,7 +64,8 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: !ref <test_batch_size>
-# Model parameters
+####################### Model Parameters #######################################
+
activation: !name:torch.nn.LeakyReLU
dnn_layers: 2
dnn_neurons: 1024
@@ -74,76 +76,12 @@ ctc_neurons: 29
output_neurons: 29 # Characters size, index(blank/eos/bos) = 0
blank_index: 0
-# Decoding parameters
-test_beam_search:
- beam_size: 200
- topk: 1
- blank_index: !ref <blank_index>
- space_token: ' ' # make sure this is the same as the one used in the tokenizer
- beam_prune_logp: -10.0
- token_prune_min_logp: -5
- prune_history: True
- alpha: 0.5
- beta: 1.5
- # can be downloaded from here https://www.openslr.org/11/ or trained with kenLM
- # It can either be a .bin or .arpa ; note: .arpa is much slower at loading
- # If you don't want to use an LM, comment it out or set it to null
- kenlm_model_path: null
-
-
#
# Functions and classes
#
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
-speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
- orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
-
-# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
-drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
-
-# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
-drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
-
-# Augmenter: Combines previously defined augmentations to perform data augmentation
-wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
- min_augmentations: 4
- max_augmentations: 4
- augment_prob: 1.0
- augmentations: [
- !ref <speed_perturb>,
- !ref <drop_freq>,
- !ref <drop_chunk>]
-
enc: !new:speechbrain.lobes.models.VanillaNN.VanillaNN
input_shape: [null, null, 1024]
@@ -214,8 +152,60 @@ lr_annealing_wav2vec: !new:speechbrain.nnet.schedulers.NewBobScheduler
annealing_factor: 0.9
patient: 0
+############################## Decoding ########################################
+
+test_beam_search:
+ beam_size: 200
+ topk: 1
+ blank_index: !ref <blank_index>
+ space_token: ' ' # make sure this is the same as the one used in the tokenizer
+ beam_prune_logp: -10.0
+ token_prune_min_logp: -5
+ prune_history: True
+ alpha: 0.5
+ beta: 1.5
+ # can be downloaded from here https://www.openslr.org/11/ or trained with kenLM
+ # It can either be a .bin or .arpa ; note: .arpa is much slower at loading
+ # If you don't want to use an LM, comment it out or set it to null
+ kenlm_model_path: null
+
label_encoder: !new:speechbrain.dataio.encoder.CTCTextEncoder
+############################## Augmentations ###################################
+
+# Speed perturbation
+speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
+ orig_freq: !ref <sample_rate>
+ speeds: [95, 100, 105]
+
+# Frequency drop: randomly drops a number of frequency bands to zero.
+drop_freq: !new:speechbrain.augment.time_domain.DropFreq
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
+
+# Time drop: randomly drops a number of temporal chunks.
+drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
+
+# Augmenter: Combines previously defined augmentations to perform data augmentation
+wav_augment: !new:speechbrain.augment.augmenter.Augmenter
+ concat_original: True
+ min_augmentations: 4
+ max_augmentations: 4
+ augment_prob: 1.0
+ augmentations: [
+ !ref <speed_perturb>,
+ !ref <drop_freq>,
+ !ref <drop_chunk>]
+
+############################## Logging and Pretrainer ##########################
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
recoverables:
diff --git a/recipes/LibriSpeech/ASR/CTC/hparams/downsampled/train_hf_wavlm_signal_downsampling.yaml b/recipes/LibriSpeech/ASR/CTC/hparams/downsampled/train_hf_wavlm_signal_downsampling.yaml
index 6c0e7207d..d0daf5b77 100644
--- a/recipes/LibriSpeech/ASR/CTC/hparams/downsampled/train_hf_wavlm_signal_downsampling.yaml
+++ b/recipes/LibriSpeech/ASR/CTC/hparams/downsampled/train_hf_wavlm_signal_downsampling.yaml
@@ -33,7 +33,8 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 1
lr: 0.9
lr_wav2vec: 0.0001
@@ -61,7 +62,8 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: !ref <test_batch_size>
-# Model parameters
+####################### Model Parameters #######################################
+
activation: !name:torch.nn.LeakyReLU
dnn_layers: 2
dnn_neurons: 1024
@@ -72,75 +74,12 @@ ctc_neurons: 58 # Twice bigger than the number of characters for upsampling
output_neurons: 29 # Characters size, index(blank/eos/bos) = 0
blank_index: 0
-# Decoding parameters
-test_beam_search:
- beam_size: 200
- topk: 1
- blank_index: !ref <blank_index>
- space_token: ' ' # make sure this is the same as the one used in the tokenizer
- beam_prune_logp: -10.0
- token_prune_min_logp: -5
- prune_history: True
- alpha: 0.5
- beta: 1.5
- # can be downloaded from here https://www.openslr.org/11/ or trained with kenLM
- # It can either be a .bin or .arpa ; note: .arpa is much slower at loading
- # If you don't want to use an LM, comment it out or set it to null
- kenlm_model_path: null
-
#
# Functions and classes
#
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
-speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
- orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
-
-# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
-drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
-
-# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
-drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
-
-# Augmenter: Combines previously defined augmentations to perform data augmentation
-wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
- min_augmentations: 4
- max_augmentations: 4
- augment_prob: 1.0
- augmentations: [
- !ref <speed_perturb>,
- !ref <drop_freq>,
- !ref <drop_chunk>]
-
enc: !new:speechbrain.lobes.models.VanillaNN.VanillaNN
input_shape: [null, null, 1024]
@@ -210,8 +149,58 @@ lr_annealing_wav2vec: !new:speechbrain.nnet.schedulers.NewBobScheduler
annealing_factor: 0.9
patient: 0
+############################## Decoding ########################################
+
+test_beam_search:
+ beam_size: 200
+ topk: 1
+ blank_index: !ref <blank_index>
+ space_token: ' ' # make sure this is the same as the one used in the tokenizer
+ beam_prune_logp: -10.0
+ token_prune_min_logp: -5
+ prune_history: True
+ alpha: 0.5
+ beta: 1.5
+ # can be downloaded from here https://www.openslr.org/11/ or trained with kenLM
+ # It can either be a .bin or .arpa ; note: .arpa is much slower at loading
+ # If you don't want to use an LM, comment it out or set it to null
+ kenlm_model_path: null
+
label_encoder: !new:speechbrain.dataio.encoder.CTCTextEncoder
+############################## Augmentations ###################################
+
+# Speed perturbation
+speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
+ orig_freq: !ref <sample_rate>
+ speeds: [95, 100, 105]
+
+# Frequency drop: randomly drops a number of frequency bands to zero.
+drop_freq: !new:speechbrain.augment.time_domain.DropFreq
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
+
+# Time drop: randomly drops a number of temporal chunks.
+drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
+
+# Augmenter: Combines previously defined augmentations to perform data augmentation
+wav_augment: !new:speechbrain.augment.augmenter.Augmenter
+ concat_original: True
+ min_augmentations: 4
+ max_augmentations: 4
+ augment_prob: 1.0
+ augmentations: [
+ !ref <speed_perturb>,
+ !ref <drop_freq>,
+ !ref <drop_chunk>]
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
recoverables:
diff --git a/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_wav2vec.yaml b/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_wav2vec.yaml
index 2d91909f2..1d860a29f 100644
--- a/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_wav2vec.yaml
+++ b/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_wav2vec.yaml
@@ -32,7 +32,8 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 1
lr: 0.9
lr_wav2vec: 0.0001
@@ -56,7 +57,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: !ref <test_batch_size>
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dnn_layers: 2
dnn_neurons: 1024
@@ -66,75 +67,14 @@ freeze_wav2vec: True
output_neurons: 29 # BPE size, index(blank/eos/bos) = 0
blank_index: 0
-# Decoding parameters
-test_beam_search:
- beam_size: 143
- topk: 1
- blank_index: !ref <blank_index>
- space_token: ' ' # make sure this is the same as the one used in the tokenizer
- beam_prune_logp: -12.0
- token_prune_min_logp: -1.2
- prune_history: True
- alpha: 0.8
- beta: 1.2
- # can be downloaded from here https://www.openslr.org/11/ or trained with kenLM
- # It can either be a .bin or .arpa ; note: .arpa is much slower at loading
- # If you don't want to use an LM, comment it out or set it to null
- kenlm_model_path: null
-
#
# Functions and classes
#
-epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
- limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
-speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
- orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
-
-# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
-drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
-
-# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
-drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
-
-# Augmenter: Combines previously defined augmentations to perform data augmentation
-wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
- min_augmentations: 4
- max_augmentations: 4
- augment_prob: 1.0
- augmentations: [
- !ref <speed_perturb>,
- !ref <drop_freq>,
- !ref <drop_chunk>]
+label_encoder: !new:speechbrain.dataio.encoder.CTCTextEncoder
+epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
+ limit: !ref <number_of_epochs>
enc: !new:speechbrain.lobes.models.VanillaNN.VanillaNN
input_shape: [null, null, 1024]
@@ -198,7 +138,58 @@ lr_annealing_wav2vec: !new:speechbrain.nnet.schedulers.NewBobScheduler
annealing_factor: 0.9
patient: 0
-label_encoder: !new:speechbrain.dataio.encoder.CTCTextEncoder
+############################## Augmentations ###################################
+
+# Speed perturbation
+speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
+ orig_freq: !ref <sample_rate>
+ speeds: [95, 100, 105]
+
+# Frequency drop: randomly drops a number of frequency bands to zero.
+drop_freq: !new:speechbrain.augment.time_domain.DropFreq
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
+
+# Time drop: randomly drops a number of temporal chunks.
+drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
+
+# Augmenter: Combines previously defined augmentations to perform data augmentation
+wav_augment: !new:speechbrain.augment.augmenter.Augmenter
+ concat_original: True
+ min_augmentations: 4
+ max_augmentations: 4
+ augment_prob: 1.0
+ augmentations: [
+ !ref <speed_perturb>,
+ !ref <drop_freq>,
+ !ref <drop_chunk>]
+
+############################## Decoding ########################################
+
+# Decoding parameters
+test_beam_search:
+ beam_size: 143
+ topk: 1
+ blank_index: !ref <blank_index>
+ space_token: ' ' # make sure this is the same as the one used in the tokenizer
+ beam_prune_logp: -12.0
+ token_prune_min_logp: -1.2
+ prune_history: True
+ alpha: 0.8
+ beta: 1.2
+ # can be downloaded from here https://www.openslr.org/11/ or trained with kenLM
+ # It can either be a .bin or .arpa ; note: .arpa is much slower at loading
+ # If you don't want to use an LM, comment it out or set it to null
+ kenlm_model_path: null
+
+############################## Logging and Pretrainer ##########################
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
diff --git a/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_wav2vec_rnn_rescoring.yaml b/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_wav2vec_rnn_rescoring.yaml
index 01a31cdd8..c946b0243 100644
--- a/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_wav2vec_rnn_rescoring.yaml
+++ b/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_wav2vec_rnn_rescoring.yaml
@@ -32,7 +32,8 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 1
lr: 0.9
lr_wav2vec: 0.0001
@@ -56,7 +57,8 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: !ref <test_batch_size>
-# Model parameters
+####################### Model Parameters #######################################
+
activation: !name:torch.nn.LeakyReLU
dnn_layers: 2
dnn_neurons: 1024
@@ -68,49 +70,6 @@ output_neurons: 29 # BPE size, index(blank/eos/bos) = 0
pretrained_lm_tokenizer_path: speechbrain/asr-crdnn-rnnlm-librispeech
-# This is the RNNLM that is used according to the Huggingface repository
-# NB: It has to match the pre-trained RNNLM!!
-lm_model: !new:speechbrain.lobes.models.RNNLM.RNNLM
- output_neurons: 1000
- embedding_dim: 128
- activation: !name:torch.nn.LeakyReLU
- dropout: 0.0
- rnn_layers: 2
- rnn_neurons: 2048
- dnn_blocks: 1
- dnn_neurons: 512
- return_hidden: True # For inference
-
-tokenizer: !new:sentencepiece.SentencePieceProcessor
-
-# Decoding parameters
-lm_weight: 0.5
-blank_index: 0
-# topk is the number of hypotheses that will be rescored in the rescorer
-# lowering this value might decrease the wer, but will increase speed.
-
-test_beam_search:
- beam_size: 20
- topk: 20
- blank_index: !ref <blank_index>
- space_token: ' ' # make sure this is the same as the one used in the tokenizer
- beam_prune_logp: -12.0
- token_prune_min_logp: -12.0
- prune_history: False
- alpha: 0.8
- beta: 1.2
-
-rnnlm: !new:speechbrain.decoders.scorer.RNNLMRescorer
- language_model: !ref <lm_model>
- tokenizer: !ref <tokenizer>
- bos_index: 0
- eos_index: 0
- pad_index: 0
-
-rescorer: !new:speechbrain.decoders.scorer.RescorerBuilder
- rescorers: [!ref <rnnlm>]
- weights:
- rnnlm: !ref <lm_weight>
#
# Functions and classes
@@ -118,53 +77,6 @@ rescorer: !new:speechbrain.decoders.scorer.RescorerBuilder
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
-speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
- orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
-
-# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
-drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
-
-# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
-drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
-
-# Augmenter: Combines previously defined augmentations to perform data augmentation
-wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
- min_augmentations: 4
- max_augmentations: 4
- augment_prob: 1.0
- augmentations: [
- !ref <speed_perturb>,
- !ref <drop_freq>,
- !ref <drop_chunk>]
-
enc: !new:speechbrain.lobes.models.VanillaNN.VanillaNN
input_shape: [null, null, 1024]
@@ -230,6 +142,84 @@ lr_annealing_wav2vec: !new:speechbrain.nnet.schedulers.NewBobScheduler
label_encoder: !new:speechbrain.dataio.encoder.CTCTextEncoder
+# This is the RNNLM that is used according to the Huggingface repository
+# NB: It has to match the pre-trained RNNLM!!
+lm_model: !new:speechbrain.lobes.models.RNNLM.RNNLM
+ output_neurons: 1000
+ embedding_dim: 128
+ activation: !name:torch.nn.LeakyReLU
+ dropout: 0.0
+ rnn_layers: 2
+ rnn_neurons: 2048
+ dnn_blocks: 1
+ dnn_neurons: 512
+ return_hidden: True # For inference
+
+
+tokenizer: !new:sentencepiece.SentencePieceProcessor
+
+############################## Decoding ########################################
+
+# topk is the number of hypotheses that will be rescored in the rescorer
+# lowering this value might decrease the wer, but will increase speed.
+test_beam_search:
+ beam_size: 20
+ topk: 20
+ blank_index: !ref <blank_index>
+ space_token: ' ' # make sure this is the same as the one used in the tokenizer
+ beam_prune_logp: -12.0
+ token_prune_min_logp: -12.0
+ prune_history: False
+ alpha: 0.8
+ beta: 1.2
+
+rnnlm: !new:speechbrain.decoders.scorer.RNNLMRescorer
+ language_model: !ref <lm_model>
+ tokenizer: !ref <tokenizer>
+ bos_index: 0
+ eos_index: 0
+ pad_index: 0
+
+rescorer: !new:speechbrain.decoders.scorer.RescorerBuilder
+ rescorers: [!ref <rnnlm>]
+ weights:
+ rnnlm: !ref <lm_weight>
+
+############################## Augmentations ###################################
+
+# Speed perturbation
+speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
+ orig_freq: !ref <sample_rate>
+ speeds: [95, 100, 105]
+
+# Frequency drop: randomly drops a number of frequency bands to zero.
+drop_freq: !new:speechbrain.augment.time_domain.DropFreq
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
+
+# Time drop: randomly drops a number of temporal chunks.
+drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
+
+# Augmenter: Combines previously defined augmentations to perform data augmentation
+wav_augment: !new:speechbrain.augment.augmenter.Augmenter
+ concat_original: True
+ min_augmentations: 4
+ max_augmentations: 4
+ augment_prob: 1.0
+ augmentations: [
+ !ref <speed_perturb>,
+ !ref <drop_freq>,
+ !ref <drop_chunk>]
+
+############################## Logging and Pretrainer ##########################
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
recoverables:
diff --git a/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_wav2vec_transformer_rescoring.yaml b/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_wav2vec_transformer_rescoring.yaml
index 724c3bf1e..d806b20cf 100644
--- a/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_wav2vec_transformer_rescoring.yaml
+++ b/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_wav2vec_transformer_rescoring.yaml
@@ -32,7 +32,8 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 1
lr: 0.9
lr_wav2vec: 0.0001
@@ -56,7 +57,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: !ref <test_batch_size>
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dnn_layers: 2
dnn_neurons: 1024
@@ -88,30 +89,6 @@ tokenizer: !new:sentencepiece.SentencePieceProcessor
# Decoding parameters
lm_weight: 0.5
blank_index: 0
-# topk is the number of hypotheses that will be rescored in the rescorer
-# lowering this value might decrease the wer, but will increase speed.
-test_beam_search:
- beam_size: 20
- topk: 20
- blank_index: !ref <blank_index>
- space_token: ' ' # make sure this is the same as the one used in the tokenizer
- beam_prune_logp: -12.0
- token_prune_min_logp: -12.0
- prune_history: False
- alpha: 0.8
- beta: 1.2
-
-transformerlm: !new:speechbrain.decoders.scorer.TransformerLMRescorer
- language_model: !ref <lm_model>
- tokenizer: !ref <tokenizer>
- pad_index: 0
- bos_index: 1
- eos_index: 2
-
-rescorer: !new:speechbrain.decoders.scorer.RescorerBuilder
- rescorers: [!ref <transformerlm>]
- weights:
- transformerlm: !ref <lm_weight>
#
# Functions and classes
@@ -119,53 +96,6 @@ rescorer: !new:speechbrain.decoders.scorer.RescorerBuilder
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
-speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
- orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
-
-# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
-drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
-
-# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
-drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
-
-# Augmenter: Combines previously defined augmentations to perform data augmentation
-wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
- min_augmentations: 4
- max_augmentations: 4
- augment_prob: 1.0
- augmentations: [
- !ref <speed_perturb>,
- !ref <drop_freq>,
- !ref <drop_chunk>]
-
enc: !new:speechbrain.lobes.models.VanillaNN.VanillaNN
input_shape: [null, null, 1024]
@@ -229,8 +159,68 @@ lr_annealing_wav2vec: !new:speechbrain.nnet.schedulers.NewBobScheduler
annealing_factor: 0.9
patient: 0
+############################## Augmentations ###################################
+
+# Speed perturbation
+speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
+ orig_freq: !ref <sample_rate>
+ speeds: [95, 100, 105]
+
+# Frequency drop: randomly drops a number of frequency bands to zero.
+drop_freq: !new:speechbrain.augment.time_domain.DropFreq
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
+
+# Time drop: randomly drops a number of temporal chunks.
+drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
+
+# Augmenter: Combines previously defined augmentations to perform data augmentation
+wav_augment: !new:speechbrain.augment.augmenter.Augmenter
+ concat_original: True
+ min_augmentations: 4
+ max_augmentations: 4
+ augment_prob: 1.0
+ augmentations: [
+ !ref <speed_perturb>,
+ !ref <drop_freq>,
+ !ref <drop_chunk>]
+
label_encoder: !new:speechbrain.dataio.encoder.CTCTextEncoder
+############################## Decoding ########################################
+
+# topk is the number of hypotheses that will be rescored in the rescorer
+# lowering this value might decrease the wer, but will increase speed.
+test_beam_search:
+ beam_size: 20
+ topk: 20
+ blank_index: !ref <blank_index>
+ space_token: ' ' # make sure this is the same as the one used in the tokenizer
+ beam_prune_logp: -12.0
+ token_prune_min_logp: -12.0
+ prune_history: False
+ alpha: 0.8
+ beta: 1.2
+
+transformerlm: !new:speechbrain.decoders.scorer.TransformerLMRescorer
+ language_model: !ref <lm_model>
+ tokenizer: !ref <tokenizer>
+ pad_index: 0
+ bos_index: 1
+ eos_index: 2
+
+rescorer: !new:speechbrain.decoders.scorer.RescorerBuilder
+ rescorers: [!ref <transformerlm>]
+ weights:
+ transformerlm: !ref <lm_weight>
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
recoverables:
diff --git a/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_whisper_encoder.yaml b/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_whisper_encoder.yaml
index 735b29db9..ba20bf2ac 100644
--- a/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_whisper_encoder.yaml
+++ b/recipes/LibriSpeech/ASR/CTC/hparams/train_hf_whisper_encoder.yaml
@@ -31,7 +31,8 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 15
warmup_steps: 1000 # We freeze whisper for 1000 steps to let the CTC adapt
lr: 0.0008
@@ -61,7 +62,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: !ref <test_batch_size>
-# Model parameters
+####################### Model Parameters #######################################
dnn_neurons: 1024
freeze_whisper: False
whisper_output_dim: 512
@@ -71,74 +72,12 @@ whisper_output_dim: 512
output_neurons: 29 # BPE size, index(blank/eos/bos) = 0
blank_index: 0
-# Decoding parameters
-test_beam_search:
- beam_size: 143
- topk: 1
- blank_index: !ref <blank_index>
- space_token: ' ' # make sure this is the same as the one used in the tokenizer
- beam_prune_logp: -12.0
- token_prune_min_logp: -1.2
- prune_history: True
- alpha: 0.8
- beta: 1.2
- # can be downloaded from here https://www.openslr.org/11/ or trained with kenLM
- # It can either be a .bin or .arpa ; note: .arpa is much slower at loading
- # If you don't want to use an LM, comment it out or set it to null
- kenlm_model_path: null
#
# Functions and classes
#
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
-speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
- orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
-
-# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
-drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
-
-# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
-drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
-
-# Augmenter: Combines previously defined augmentations to perform data augmentation
-wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
- min_augmentations: 4
- max_augmentations: 4
- augment_prob: 1.0
- augmentations: [
- !ref <speed_perturb>,
- !ref <drop_freq>,
- !ref <drop_chunk>]
-
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <whisper_output_dim>]
@@ -204,6 +143,57 @@ lr_annealing_whisper: !new:speechbrain.nnet.schedulers.NewBobScheduler
annealing_factor: 0.75
patient: 0
+############################## Decoding ########################################
+
+test_beam_search:
+ beam_size: 143
+ topk: 1
+ blank_index: !ref <blank_index>
+ space_token: ' ' # make sure this is the same as the one used in the tokenizer
+ beam_prune_logp: -12.0
+ token_prune_min_logp: -1.2
+ prune_history: True
+ alpha: 0.8
+ beta: 1.2
+ # can be downloaded from here https://www.openslr.org/11/ or trained with kenLM
+ # It can either be a .bin or .arpa ; note: .arpa is much slower at loading
+ # If you don't want to use an LM, comment it out or set it to null
+ kenlm_model_path: null
+
+############################## Augmentations ###################################
+
+# Speed perturbation
+speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
+ orig_freq: !ref <sample_rate>
+ speeds: [95, 100, 105]
+
+# Frequency drop: randomly drops a number of frequency bands to zero.
+drop_freq: !new:speechbrain.augment.time_domain.DropFreq
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
+
+# Time drop: randomly drops a number of temporal chunks.
+drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
+
+# Augmenter: Combines previously defined augmentations to perform data augmentation
+wav_augment: !new:speechbrain.augment.augmenter.Augmenter
+ concat_original: True
+ min_augmentations: 4
+ max_augmentations: 4
+ augment_prob: 1.0
+ augmentations: [
+ !ref <speed_perturb>,
+ !ref <drop_freq>,
+ !ref <drop_chunk>]
+
+############################## Logging and Pretrainer ##########################
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
diff --git a/recipes/LibriSpeech/ASR/CTC/hparams/train_sb_wav2vec.yaml b/recipes/LibriSpeech/ASR/CTC/hparams/train_sb_wav2vec.yaml
index cf9cf7ec8..1b281b35c 100644
--- a/recipes/LibriSpeech/ASR/CTC/hparams/train_sb_wav2vec.yaml
+++ b/recipes/LibriSpeech/ASR/CTC/hparams/train_sb_wav2vec.yaml
@@ -33,7 +33,7 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
lr: 0.0003
lr_wav2vec: 0.00005
@@ -58,7 +58,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: !ref <test_batch_size>
-# Model parameters
+####################### Model Parameters #######################################
dnn_activation: !new:torch.nn.LeakyReLU
dnn_neurons: 1280
dnn_dropout: 0.15
@@ -68,75 +68,12 @@ freeze_wav2vec: False
output_neurons: 29 # BPE size, index(blank/eos/bos) = 0
blank_index: 0
-# Decoding parameters
-test_beam_search:
- beam_size: 200
- topk: 1
- blank_index: !ref <blank_index>
- space_token: ' ' # make sure this is the same as the one used in the tokenizer
- beam_prune_logp: -10.0
- token_prune_min_logp: -5.0
- prune_history: True
- alpha: 0.8
- beta: 1.2
- # can be downloaded from here https://www.openslr.org/11/ or trained with kenLM
- # It can either be a .bin or .arpa ; note: .arpa is much slower at loading
- # If you don't want to use an LM, comment it out or set it to null
- kenlm_model_path: null
-
#
# Functions and classes
#
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
-speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
- orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
-
-# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
-drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
-
-# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
-drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
-
-# Augmenter: Combines previously defined augmentations to perform data augmentation
-wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
- min_augmentations: 4
- max_augmentations: 4
- augment_prob: 1.0
- augmentations: [
- !ref <speed_perturb>,
- !ref <drop_freq>,
- !ref <drop_chunk>]
-
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
@@ -227,6 +164,58 @@ lr_annealing_wav2vec: !new:speechbrain.nnet.schedulers.NewBobScheduler
annealing_factor: 0.7
patient: 0
+############################## Decoding ########################################
+
+test_beam_search:
+ beam_size: 200
+ topk: 1
+ blank_index: !ref <blank_index>
+ space_token: ' ' # make sure this is the same as the one used in the tokenizer
+ beam_prune_logp: -10.0
+ token_prune_min_logp: -5.0
+ prune_history: True
+ alpha: 0.8
+ beta: 1.2
+ # can be downloaded from here https://www.openslr.org/11/ or trained with kenLM
+ # It can either be a .bin or .arpa ; note: .arpa is much slower at loading
+ # If you don't want to use an LM, comment it out or set it to null
+ kenlm_model_path: null
+
+############################## Augmentations ###################################
+
+# Speed perturbation
+speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
+ orig_freq: !ref <sample_rate>
+ speeds: [95, 100, 105]
+
+# Frequency drop: randomly drops a number of frequency bands to zero.
+drop_freq: !new:speechbrain.augment.time_domain.DropFreq
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
+
+# Time drop: randomly drops a number of temporal chunks.
+drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
+
+# Augmenter: Combines previously defined augmentations to perform data augmentation
+wav_augment: !new:speechbrain.augment.augmenter.Augmenter
+ concat_original: True
+ min_augmentations: 4
+ max_augmentations: 4
+ augment_prob: 1.0
+ augmentations: [
+ !ref <speed_perturb>,
+ !ref <drop_freq>,
+ !ref <drop_chunk>]
+
+############################## Logging and Pretrainer ##########################
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
recoverables:
diff --git a/recipes/LibriSpeech/ASR/CTC/train_with_wav2vec.py b/recipes/LibriSpeech/ASR/CTC/train_with_wav2vec.py
index f2b6373b2..1f4ccdd2c 100644
--- a/recipes/LibriSpeech/ASR/CTC/train_with_wav2vec.py
+++ b/recipes/LibriSpeech/ASR/CTC/train_with_wav2vec.py
@@ -101,10 +101,15 @@ class ASR(sb.Brain):
ids = batch.id
tokens, tokens_lens = batch.tokens
- # Label Augmentation
+ # Labels must be extended if parallel augmentation or concatenated
+ # augmentation was performed on the input (increasing the time dimension)
if stage == sb.Stage.TRAIN and hasattr(self.hparams, "wav_augment"):
- tokens = self.hparams.wav_augment.replicate_labels(tokens)
- tokens_lens = self.hparams.wav_augment.replicate_labels(tokens_lens)
+ (
+ tokens,
+ tokens_lens,
+ ) = self.hparams.wav_augment.replicate_multiple_labels(
+ tokens, tokens_lens
+ )
loss_ctc = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
loss = loss_ctc
diff --git a/recipes/LibriSpeech/ASR/CTC/train_with_whisper.py b/recipes/LibriSpeech/ASR/CTC/train_with_whisper.py
index e3db36334..d575265e8 100644
--- a/recipes/LibriSpeech/ASR/CTC/train_with_whisper.py
+++ b/recipes/LibriSpeech/ASR/CTC/train_with_whisper.py
@@ -72,10 +72,15 @@ class ASR(sb.Brain):
ids = batch.id
tokens, tokens_lens = batch.tokens
- # Label Augmentation
+ # Labels must be extended if parallel augmentation or concatenated
+ # augmentation was performed on the input (increasing the time dimension)
if stage == sb.Stage.TRAIN and hasattr(self.hparams, "wav_augment"):
- tokens = self.hparams.wav_augment.replicate_labels(tokens)
- tokens_lens = self.hparams.wav_augment.replicate_labels(tokens_lens)
+ (
+ tokens,
+ tokens_lens,
+ ) = self.hparams.wav_augment.replicate_multiple_labels(
+ tokens, tokens_lens
+ )
loss_ctc = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
loss = loss_ctc
diff --git a/recipes/LibriSpeech/ASR/seq2seq/hparams/train_BPE_1000.yaml b/recipes/LibriSpeech/ASR/seq2seq/hparams/train_BPE_1000.yaml
index e38b545fb..3d0aaa200 100644
--- a/recipes/LibriSpeech/ASR/seq2seq/hparams/train_BPE_1000.yaml
+++ b/recipes/LibriSpeech/ASR/seq2seq/hparams/train_BPE_1000.yaml
@@ -44,7 +44,8 @@ data_folder_noise: !ref <data_folder>/noise # The noisy sequencies for data augm
NOISE_DATASET_URL: https://www.dropbox.com/scl/fi/a09pj97s5ifan81dqhi4n/noises.zip?rlkey=j8b0n9kdjdr32o1f06t0cw5b7&dl=1
noise_annotation: !ref <save_folder>/noise.csv #The data manifest files are created by the data preparation script
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 15
number_of_ctc_epochs: 5
batch_size: 8
@@ -89,7 +90,8 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-# Model parameters
+####################### Model Parameters #######################################
+
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 2
@@ -128,74 +130,6 @@ coverage_penalty: 1.5
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Download and prepare the dataset of noisy sequences for augmentation
-prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
- URL: !ref <NOISE_DATASET_URL>
- dest_folder: !ref <data_folder_noise>
- ext: wav
- csv_file: !ref <noise_annotation>
-
-
-# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
-add_noise: !new:speechbrain.augment.time_domain.AddNoise
- csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
- noise_sample_rate: !ref <sample_rate>
- clean_sample_rate: !ref <sample_rate>
- num_workers: !ref <num_workers>
-
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
-speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
- orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
-
-# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
-drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
-
-# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
-drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
-
-# Augmenter: Combines previously defined augmentations to perform data augmentation
-wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
- min_augmentations: 4
- max_augmentations: 4
- augment_prob: 1.0
- augmentations: [
- !ref <add_noise>,
- !ref <speed_perturb>,
- !ref <drop_freq>,
- !ref <drop_chunk>]
-
normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
@@ -288,7 +222,8 @@ modules:
model: !new:torch.nn.ModuleList
- [!ref <enc>, !ref <emb>, !ref <dec>, !ref <ctc_lin>, !ref <seq_lin>]
-# Scorer
+############################## Decoding & optimiser ############################
+
coverage_scorer: !new:speechbrain.decoders.scorer.CoverageScorer
vocab_size: !ref <output_neurons>
@@ -339,6 +274,57 @@ lr_annealing: !new:speechbrain.nnet.schedulers.NewBobScheduler
annealing_factor: 0.8
patient: 0
+############################## Augmentations ###################################
+
+prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
+ URL: !ref <NOISE_DATASET_URL>
+ dest_folder: !ref <data_folder_noise>
+ ext: wav
+ csv_file: !ref <noise_annotation>
+
+# Add noise to input signal
+add_noise: !new:speechbrain.augment.time_domain.AddNoise
+ csv_file: !ref <noise_annotation>
+ snr_low: 0
+ snr_high: 15
+ noise_sample_rate: !ref <sample_rate>
+ clean_sample_rate: !ref <sample_rate>
+ num_workers: !ref <num_workers>
+
+# Speed perturbation
+speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
+ orig_freq: !ref <sample_rate>
+ speeds: [95, 100, 105]
+
+# Frequency drop: randomly drops a number of frequency bands to zero.
+drop_freq: !new:speechbrain.augment.time_domain.DropFreq
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
+
+# Time drop: randomly drops a number of temporal chunks.
+drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
+
+# Augmenter: Combines previously defined augmentations to perform data augmentation
+wav_augment: !new:speechbrain.augment.augmenter.Augmenter
+ concat_original: True
+ min_augmentations: 4
+ max_augmentations: 4
+ augment_prob: 1.0
+ augmentations: [
+ !ref <add_noise>,
+ !ref <speed_perturb>,
+ !ref <drop_freq>,
+ !ref <drop_chunk>]
+
+############################## Logging and Pretrainer ##########################
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
recoverables:
diff --git a/recipes/LibriSpeech/ASR/seq2seq/hparams/train_BPE_1000_sligru.yaml b/recipes/LibriSpeech/ASR/seq2seq/hparams/train_BPE_1000_sligru.yaml
index 164f1ffe7..355c49d36 100644
--- a/recipes/LibriSpeech/ASR/seq2seq/hparams/train_BPE_1000_sligru.yaml
+++ b/recipes/LibriSpeech/ASR/seq2seq/hparams/train_BPE_1000_sligru.yaml
@@ -44,7 +44,8 @@ data_folder_noise: !ref <data_folder>/noise # The noisy sequencies for data augm
NOISE_DATASET_URL: https://www.dropbox.com/scl/fi/a09pj97s5ifan81dqhi4n/noises.zip?rlkey=j8b0n9kdjdr32o1f06t0cw5b7&dl=1
noise_annotation: !ref <save_folder>/noise.csv #The data manifest files are created by the data preparation script
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 15
number_of_ctc_epochs: 15
batch_size: 24
@@ -89,7 +90,8 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-# Model parameters
+####################### Model Parameters #######################################
+
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 2
@@ -126,75 +128,6 @@ temperature_lm: 1.25
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Download and prepare the dataset of noisy sequences for augmentation
-prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
- URL: !ref <NOISE_DATASET_URL>
- dest_folder: !ref <data_folder_noise>
- ext: wav
- csv_file: !ref <noise_annotation>
-
-
-# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
-add_noise: !new:speechbrain.augment.time_domain.AddNoise
- csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
- noise_sample_rate: !ref <sample_rate>
- clean_sample_rate: !ref <sample_rate>
- num_workers: !ref <num_workers>
-
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
-speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
- orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
-
-# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
-drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
-
-# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
-drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
-
-# Augmenter: Combines previously defined augmentations to perform data augmentation
-wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
- min_augmentations: 4
- max_augmentations: 4
- augment_prob: 1.0
- augmentations: [
- !ref <add_noise>,
- !ref <speed_perturb>,
- !ref <drop_freq>,
- !ref <drop_chunk>]
-
-
normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
@@ -286,7 +219,8 @@ modules:
model: !new:torch.nn.ModuleList
- [!ref <enc>, !ref <emb>, !ref <dec>, !ref <ctc_lin>, !ref <seq_lin>]
-# Scorer
+############################## Decoding & optimiser ############################
+
coverage_scorer: !new:speechbrain.decoders.scorer.CoverageScorer
vocab_size: !ref <output_neurons>
@@ -337,6 +271,57 @@ lr_annealing: !new:speechbrain.nnet.schedulers.NewBobScheduler
annealing_factor: 0.8
patient: 0
+############################## Augmentations ###################################
+
+prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
+ URL: !ref <NOISE_DATASET_URL>
+ dest_folder: !ref <data_folder_noise>
+ ext: wav
+ csv_file: !ref <noise_annotation>
+
+# Add noise to input signal
+add_noise: !new:speechbrain.augment.time_domain.AddNoise
+ csv_file: !ref <noise_annotation>
+ snr_low: 0
+ snr_high: 15
+ noise_sample_rate: !ref <sample_rate>
+ clean_sample_rate: !ref <sample_rate>
+ num_workers: !ref <num_workers>
+
+# Speed perturbation
+speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
+ orig_freq: !ref <sample_rate>
+ speeds: [95, 100, 105]
+
+# Frequency drop: randomly drops a number of frequency bands to zero.
+drop_freq: !new:speechbrain.augment.time_domain.DropFreq
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
+
+# Time drop: randomly drops a number of temporal chunks.
+drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
+
+# Augmenter: Combines previously defined augmentations to perform data augmentation
+wav_augment: !new:speechbrain.augment.augmenter.Augmenter
+ concat_original: True
+ min_augmentations: 4
+ max_augmentations: 4
+ augment_prob: 1.0
+ augmentations: [
+ !ref <add_noise>,
+ !ref <speed_perturb>,
+ !ref <drop_freq>,
+ !ref <drop_chunk>]
+
+############################## Logging and Pretrainer ##########################
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
recoverables:
diff --git a/recipes/LibriSpeech/ASR/seq2seq/hparams/train_BPE_5000.yaml b/recipes/LibriSpeech/ASR/seq2seq/hparams/train_BPE_5000.yaml
index cc0647562..3046dfea8 100644
--- a/recipes/LibriSpeech/ASR/seq2seq/hparams/train_BPE_5000.yaml
+++ b/recipes/LibriSpeech/ASR/seq2seq/hparams/train_BPE_5000.yaml
@@ -45,7 +45,8 @@ data_folder_noise: !ref <data_folder>/noise # The noisy sequencies for data augm
NOISE_DATASET_URL: https://www.dropbox.com/scl/fi/a09pj97s5ifan81dqhi4n/noises.zip?rlkey=j8b0n9kdjdr32o1f06t0cw5b7&dl=1
noise_annotation: !ref <save_folder>/noise.csv #The data manifest files are created by the data preparation script
-# Training parameters
+####################### Training Parameters ####################################
+
number_of_epochs: 25
number_of_ctc_epochs: 25
batch_size: 8
@@ -90,7 +91,8 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-# Model parameters
+####################### Model Parameters #######################################
+
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 2
@@ -132,75 +134,6 @@ coverage_penalty: 1.5
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Download and prepare the dataset of noisy sequences for augmentation
-prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
- URL: !ref <NOISE_DATASET_URL>
- dest_folder: !ref <data_folder_noise>
- ext: wav
- csv_file: !ref <noise_annotation>
-
-
-# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
-add_noise: !new:speechbrain.augment.time_domain.AddNoise
- csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
- noise_sample_rate: !ref <sample_rate>
- clean_sample_rate: !ref <sample_rate>
- num_workers: !ref <num_workers>
-
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
-speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
- orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
-
-# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
-drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
-
-# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
-drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
-
-# Augmenter: Combines previously defined augmentations to perform data augmentation
-wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
- min_augmentations: 4
- max_augmentations: 4
- augment_prob: 1.0
- augmentations: [
- !ref <add_noise>,
- !ref <speed_perturb>,
- !ref <drop_freq>,
- !ref <drop_chunk>]
-
-
normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
@@ -294,7 +227,8 @@ modules:
model: !new:torch.nn.ModuleList
- [!ref <enc>, !ref <emb>, !ref <dec>, !ref <ctc_lin>, !ref <seq_lin>]
-# Scorer
+############################## Decoding & optimiser ############################
+
ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
eos_index: !ref <eos_index>
blank_index: !ref <blank_index>
@@ -360,6 +294,57 @@ lr_annealing: !new:speechbrain.nnet.schedulers.NewBobScheduler
annealing_factor: 0.8
patient: 0
+############################## Augmentations ###################################
+
+prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
+ URL: !ref <NOISE_DATASET_URL>
+ dest_folder: !ref <data_folder_noise>
+ ext: wav
+ csv_file: !ref <noise_annotation>
+
+# Add noise to input signal
+add_noise: !new:speechbrain.augment.time_domain.AddNoise
+ csv_file: !ref <noise_annotation>
+ snr_low: 0
+ snr_high: 15
+ noise_sample_rate: !ref <sample_rate>
+ clean_sample_rate: !ref <sample_rate>
+ num_workers: !ref <num_workers>
+
+# Speed perturbation
+speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
+ orig_freq: !ref <sample_rate>
+ speeds: [95, 100, 105]
+
+# Frequency drop: randomly drops a number of frequency bands to zero.
+drop_freq: !new:speechbrain.augment.time_domain.DropFreq
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
+
+# Time drop: randomly drops a number of temporal chunks.
+drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
+
+# Augmenter: Combines previously defined augmentations to perform data augmentation
+wav_augment: !new:speechbrain.augment.augmenter.Augmenter
+ concat_original: True
+ min_augmentations: 4
+ max_augmentations: 4
+ augment_prob: 1.0
+ augmentations: [
+ !ref <add_noise>,
+ !ref <speed_perturb>,
+ !ref <drop_freq>,
+ !ref <drop_chunk>]
+
+############################## Logging and Pretrainer ##########################
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
recoverables:
diff --git a/recipes/LibriSpeech/ASR/seq2seq/train.py b/recipes/LibriSpeech/ASR/seq2seq/train.py
index b3adaa67a..7f5351008 100644
--- a/recipes/LibriSpeech/ASR/seq2seq/train.py
+++ b/recipes/LibriSpeech/ASR/seq2seq/train.py
@@ -97,12 +97,16 @@ class ASR(sb.Brain):
tokens_eos, tokens_eos_lens = batch.tokens_eos
tokens, tokens_lens = batch.tokens
+ # Labels must be extended if parallel augmentation or concatenated
+ # augmentation was performed on the input (increasing the time dimension)
if stage == sb.Stage.TRAIN and hasattr(self.hparams, "wav_augment"):
- tokens = self.hparams.wav_augment.replicate_labels(tokens)
- tokens_lens = self.hparams.wav_augment.replicate_labels(tokens_lens)
- tokens_eos = self.hparams.wav_augment.replicate_labels(tokens_eos)
- tokens_eos_lens = self.hparams.wav_augment.replicate_labels(
- tokens_eos_lens
+ (
+ tokens,
+ tokens_lens,
+ tokens_eos,
+ tokens_eos_lens,
+ ) = self.hparams.wav_augment.replicate_multiple_labels(
+ tokens, tokens_lens, tokens_eos, tokens_eos_lens
)
loss_seq = self.hparams.seq_cost(
diff --git a/recipes/LibriSpeech/ASR/transducer/hparams/conformer_transducer.yaml b/recipes/LibriSpeech/ASR/transducer/hparams/conformer_transducer.yaml
index c7ad99c63..e9757e208 100644
--- a/recipes/LibriSpeech/ASR/transducer/hparams/conformer_transducer.yaml
+++ b/recipes/LibriSpeech/ASR/transducer/hparams/conformer_transducer.yaml
@@ -40,7 +40,8 @@ test_csv:
skip_prep: False
ckpt_interval_minutes: 5 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
+
# To make Transformers converge, the global bath size should be large enough.
# The global batch size is computed as batch_size * n_gpus * grad_accumulation_factor.
# Empirically, we found that this value should be >= 128.
@@ -120,7 +121,8 @@ dynamic_batch_sampler:
batch_ordering: random
max_batch_ex: 256
-# Model parameters
+####################### Model Parameters #######################################
+
# Transformer
d_model: 512
joint_dim: 640
@@ -164,18 +166,15 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_mels: !ref <n_mels>
win_length: !ref <win_length>
+############################## Augmentations ###################################
+
# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 1
max_augmentations: 1
augment_prob: 1.0
@@ -183,43 +182,24 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
-
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
+ replace: "zeros"
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 2
+ drop_count_high: 2
+ replace: "zeros"
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
parallel_augment: False
@@ -234,6 +214,8 @@ fea_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <freq_drop>,
!ref <time_warp>]
+############################## Models ##########################################
+
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
input_shape: (8, 10, 80)
num_blocks: 2
@@ -355,6 +337,8 @@ modules:
model: !new:torch.nn.ModuleList
- [!ref <CNN>, !ref <enc>, !ref <emb>, !ref <dec>, !ref <proj_enc>, !ref <proj_dec>, !ref <proj_ctc>, !ref <transducer_lin>]
+############################## Decoding & optimiser ############################
+
# Tokenizer initialization
tokenizer: !new:sentencepiece.SentencePieceProcessor
@@ -388,6 +372,8 @@ noam_annealing: !new:speechbrain.nnet.schedulers.NoamScheduler
lr_initial: !ref <lr>
n_warmup_steps: !ref <warmup_steps>
+############################## Logging and Pretrainer ##########################
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
recoverables:
diff --git a/recipes/LibriSpeech/ASR/transducer/train.py b/recipes/LibriSpeech/ASR/transducer/train.py
index 497912c83..84d7e05ff 100644
--- a/recipes/LibriSpeech/ASR/transducer/train.py
+++ b/recipes/LibriSpeech/ASR/transducer/train.py
@@ -155,27 +155,16 @@ class ASR(sb.Brain):
logits_transducer, wav_lens, predicted_tokens = predictions
if stage == sb.Stage.TRAIN:
- if hasattr(self.hparams, "wav_augment"):
- tokens = self.hparams.wav_augment.replicate_labels(tokens)
- token_lens = self.hparams.wav_augment.replicate_labels(
- token_lens
- )
- tokens_eos = self.hparams.wav_augment.replicate_labels(
- tokens_eos
- )
- token_eos_lens = self.hparams.wav_augment.replicate_labels(
- token_eos_lens
- )
+ # Labels must be extended if parallel augmentation or concatenated
+ # augmentation was performed on the input (increasing the time dimension)
if hasattr(self.hparams, "fea_augment"):
- tokens = self.hparams.fea_augment.replicate_labels(tokens)
- token_lens = self.hparams.fea_augment.replicate_labels(
- token_lens
- )
- tokens_eos = self.hparams.fea_augment.replicate_labels(
- tokens_eos
- )
- token_eos_lens = self.hparams.fea_augment.replicate_labels(
- token_eos_lens
+ (
+ tokens,
+ token_lens,
+ tokens_eos,
+ token_eos_lens,
+ ) = self.hparams.fea_augment.replicate_multiple_labels(
+ tokens, token_lens, tokens_eos, token_eos_lens
)
if stage == sb.Stage.TRAIN:
diff --git a/recipes/LibriSpeech/ASR/transformer/hparams/bayesspeech.yaml b/recipes/LibriSpeech/ASR/transformer/hparams/bayesspeech.yaml
index 7772517e9..2eee3646e 100644
--- a/recipes/LibriSpeech/ASR/transformer/hparams/bayesspeech.yaml
+++ b/recipes/LibriSpeech/ASR/transformer/hparams/bayesspeech.yaml
@@ -42,7 +42,8 @@ test_csv:
ckpt_interval_minutes: 30 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
+
# To make Transformers converge, the global bath size should be large enough.
# The global batch size is computed as batch_size * n_gpus * grad_accumulation_factor.
# Empirically, we found that this value should be >= 128.
@@ -116,7 +117,7 @@ test_dataloader_opts:
padding_kwargs:
value: !ref <pad_index>
-####################### Model parameters ###########################
+####################### Model Parameters #######################################
# Transformer
d_model: 512
nhead: 4
@@ -148,7 +149,7 @@ test_beam_size: 66
lm_weight: 0.60
ctc_weight_decode: 0.40
-############################## models ################################
+############################## Models ##########################################
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
input_shape: (8, 10, 80)
@@ -216,7 +217,8 @@ Adam: !name:torch.optim.Adam
eps: 0.000000001
-# Scorer
+############################## Decoding & optimiser ############################
+
ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
eos_index: !ref <eos_index>
blank_index: !ref <blank_index>
@@ -290,57 +292,34 @@ normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
update_until_epoch: 4
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 4 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 4 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "mean" # Method of dropping chunks
-
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
# Frequency Drop
-freq_drop_length_low: 10 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 20 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 4 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 4 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "mean" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 10
+ drop_length_high: 20
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -354,6 +333,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Logging and Pretrainer ##########################
+
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
save_file: !ref <train_log>
diff --git a/recipes/LibriSpeech/ASR/transformer/hparams/branchformer_large.yaml b/recipes/LibriSpeech/ASR/transformer/hparams/branchformer_large.yaml
index c3f66ec9f..02fc2eac4 100644
--- a/recipes/LibriSpeech/ASR/transformer/hparams/branchformer_large.yaml
+++ b/recipes/LibriSpeech/ASR/transformer/hparams/branchformer_large.yaml
@@ -41,9 +41,11 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+####################### Training Parameters ####################################
+
# To make Transformers converge, the global bath size should be large enough.
-# The global batch size is computed as batch_size * n_gpus * grad_accumulation_factor.
+# The global batch size is computed as batch_size * n_gpus *
+# grad_accumulation_factor.
# Empirically, we found that this value should be >= 128.
# Please, set your parameters accordingly.
number_of_epochs: 120
@@ -103,7 +105,8 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-####################### Model parameters ###########################
+####################### Model Parameters #######################################
+
# Transformer
d_model: 512
nhead: 8
@@ -131,7 +134,7 @@ test_beam_size: 66
lm_weight: 0.60
ctc_weight_decode: 0.40
-############################## models ################################
+############################## Models ##########################################
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
input_shape: (8, 10, 80)
@@ -204,7 +207,8 @@ Adam: !name:torch.optim.AdamW
eps: 0.000000001
weight_decay: !ref <weight_decay>
-# Scorer
+####################### Decoding & optimiser ###################################
+
ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
eos_index: !ref <eos_index>
blank_index: !ref <blank_index>
@@ -275,57 +279,34 @@ checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+####################### Augmentations ###########################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 4 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 4 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "mean" # Method of dropping chunks
-
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
-
-# Frequency Drop
-freq_drop_length_low: 10 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 20 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 4 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 4 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "mean" # Method of dropping chunks
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
+# Freq Drop
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 10
+ drop_length_high: 20
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -340,6 +321,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
win_length: !ref <win_length>
n_mels: !ref <n_mels>
+############################## Logging and Pretrainer ##########################
+
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
save_file: !ref <train_log>
diff --git a/recipes/LibriSpeech/ASR/transformer/hparams/conformer_large.yaml b/recipes/LibriSpeech/ASR/transformer/hparams/conformer_large.yaml
index 5d252a4b5..7cdd4c06f 100644
--- a/recipes/LibriSpeech/ASR/transformer/hparams/conformer_large.yaml
+++ b/recipes/LibriSpeech/ASR/transformer/hparams/conformer_large.yaml
@@ -41,7 +41,8 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+####################### Training Parameters ####################################
+
# To make Transformers converge, the global bath size should be large enough.
# The global batch size is computed as batch_size * n_gpus * grad_accumulation_factor.
# Empirically, we found that this value should be >= 128.
@@ -102,7 +103,8 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-####################### Model parameters ###########################
+####################### Model Parameters #######################################
+
# Transformer
d_model: 512
nhead: 8
@@ -129,7 +131,7 @@ test_beam_size: 66
lm_weight: 0.60
ctc_weight_decode: 0.40
-############################## models ################################
+############################## Models ##########################################
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
input_shape: (8, 10, 80)
@@ -200,7 +202,8 @@ Adam: !name:torch.optim.AdamW
model: !new:torch.nn.ModuleList
- [!ref <CNN>, !ref <Transformer>, !ref <seq_lin>, !ref <ctc_lin>]
-# Scorer
+####################### Decoding & optimiser ###########################
+
ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
eos_index: !ref <eos_index>
blank_index: !ref <blank_index>
@@ -271,57 +274,34 @@ checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 4 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 4 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "mean" # Method of dropping chunks
-
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
-
-# Frequency Drop
-freq_drop_length_low: 10 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 20 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 4 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 4 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "mean" # Method of dropping chunks
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
+# Freq Drop
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 10
+ drop_length_high: 20
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -336,6 +316,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_mels: !ref <n_mels>
win_length: !ref <win_length>
+############################## Logging and Pretrainer ##########################
+
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
save_file: !ref <train_log>
diff --git a/recipes/LibriSpeech/ASR/transformer/hparams/conformer_small.yaml b/recipes/LibriSpeech/ASR/transformer/hparams/conformer_small.yaml
index eddc96780..a24e6649a 100644
--- a/recipes/LibriSpeech/ASR/transformer/hparams/conformer_small.yaml
+++ b/recipes/LibriSpeech/ASR/transformer/hparams/conformer_small.yaml
@@ -40,7 +40,8 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+####################### Training Parameters ####################################
+
# To make Transformers converge, the global bath size should be large enough.
# The global batch size is computed as batch_size * n_gpus * grad_accumulation_factor.
# Empirically, we found that this value should be >= 128.
@@ -102,7 +103,8 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-####################### Model parameters ###########################
+####################### Model Parameters #######################################
+
# Transformer
d_model: 144
nhead: 4
@@ -129,7 +131,7 @@ test_beam_size: 66
lm_weight: 0.60
ctc_weight_decode: 0.40
-############################## models ################################
+############################## Models ##########################################
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
input_shape: (8, 10, 80)
@@ -200,12 +202,8 @@ Adam: !name:torch.optim.Adam
betas: (0.9, 0.98)
eps: 0.000000001
-#SGD: !name:torch.optim.SGD
-# lr: !ref <lr_sgd>
-# momentum: 0.99
-# nesterov: True
+############################## Decoding & optimiser ############################
-# Scorer
ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
eos_index: !ref <eos_index>
blank_index: !ref <blank_index>
@@ -277,57 +275,34 @@ checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 4 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 4 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "mean" # Method of dropping chunks
-
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
-
-# Frequency Drop
-freq_drop_length_low: 10 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 20 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 4 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 4 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "mean" # Method of dropping chunks
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
+# Freq Drop
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 10
+ drop_length_high: 20
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -341,6 +316,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Logging and Pretrainer ##########################
+
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
save_file: !ref <train_log>
diff --git a/recipes/LibriSpeech/ASR/transformer/hparams/hyperbranchformer_13M.yaml b/recipes/LibriSpeech/ASR/transformer/hparams/hyperbranchformer_13M.yaml
index 7b2912ec3..4b6ca718f 100644
--- a/recipes/LibriSpeech/ASR/transformer/hparams/hyperbranchformer_13M.yaml
+++ b/recipes/LibriSpeech/ASR/transformer/hparams/hyperbranchformer_13M.yaml
@@ -40,7 +40,8 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+############################## Training Parameters #############################
+
# To make Transformers converge, the global bath size should be large enough.
# The global batch size is computed as batch_size * n_gpus * grad_accumulation_factor.
# Empirically, we found that this value should be >= 128.
@@ -99,7 +100,8 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-####################### Model parameters ###########################
+####################### Model Parameters #######################################
+
# Transformer
d_model: 144
nhead: 8
@@ -133,7 +135,7 @@ test_beam_size: 66
lm_weight: 0.60
ctc_weight_decode: 0.40
-############################## models ################################
+############################## Models ##########################################
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
input_shape: (8, 10, 80)
@@ -206,7 +208,8 @@ Adam: !name:torch.optim.Adam
betas: (0.9, 0.98)
eps: 0.000000001
-# Scorer
+############################## Decoding & optimiser ############################
+
ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
eos_index: !ref <eos_index>
blank_index: !ref <blank_index>
@@ -276,57 +279,34 @@ checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 4 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 4 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "mean" # Method of dropping chunks
-
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
-
-# Frequency Drop
-freq_drop_length_low: 10 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 20 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 4 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 4 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "mean" # Method of dropping chunks
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
+# Freq Drop
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 10
+ drop_length_high: 20
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -340,6 +320,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Logging and Pretrainer ##########################
+
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
save_file: !ref <train_log>
diff --git a/recipes/LibriSpeech/ASR/transformer/hparams/hyperbranchformer_25M.yaml b/recipes/LibriSpeech/ASR/transformer/hparams/hyperbranchformer_25M.yaml
index 71d97cdb9..2e0242e31 100644
--- a/recipes/LibriSpeech/ASR/transformer/hparams/hyperbranchformer_25M.yaml
+++ b/recipes/LibriSpeech/ASR/transformer/hparams/hyperbranchformer_25M.yaml
@@ -40,7 +40,8 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+############################## Training Parameters #############################
+
# To make Transformers converge, the global bath size should be large enough.
# The global batch size is computed as batch_size * n_gpus * grad_accumulation_factor.
# Empirically, we found that this value should be >= 128.
@@ -100,7 +101,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-####################### Model parameters ###########################
+####################### Model Parameters #######################################
# Transformer
d_model: 256
nhead: 8
@@ -134,7 +135,7 @@ test_beam_size: 66
lm_weight: 0.60
ctc_weight_decode: 0.40
-############################## models ################################
+############################## Models ##########################################
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
input_shape: (8, 10, 80)
@@ -207,7 +208,8 @@ Adam: !name:torch.optim.Adam
betas: (0.9, 0.98)
eps: 0.000000001
-# Scorer
+############################## Decoding & optimiser ############################
+
ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
eos_index: !ref <eos_index>
blank_index: !ref <blank_index>
@@ -277,57 +279,34 @@ checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 4 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 4 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "mean" # Method of dropping chunks
-
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
-
-# Frequency Drop
-freq_drop_length_low: 10 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 20 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 4 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 4 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "mean" # Method of dropping chunks
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
+# Freq Drop
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 10
+ drop_length_high: 20
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -335,11 +314,14 @@ fea_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <time_drop>,
!ref <freq_drop>,
!ref <time_warp>]
+
compute_features: !new:speechbrain.lobes.features.Fbank
sample_rate: !ref <sample_rate>
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Logging and Pretrainer ##########################
+
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
save_file: !ref <train_log>
diff --git a/recipes/LibriSpeech/ASR/transformer/hparams/hyperconformer_22M.yaml b/recipes/LibriSpeech/ASR/transformer/hparams/hyperconformer_22M.yaml
index fdf65fde3..6e165ed5c 100644
--- a/recipes/LibriSpeech/ASR/transformer/hparams/hyperconformer_22M.yaml
+++ b/recipes/LibriSpeech/ASR/transformer/hparams/hyperconformer_22M.yaml
@@ -40,7 +40,8 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+####################### Training Parameters ####################################
+
# To make Transformers converge, the global bath size should be large enough.
# The global batch size is computed as batch_size * n_gpus * grad_accumulation_factor.
# Empirically, we found that this value should be >= 128.
@@ -87,6 +88,7 @@ dynamic_batch_sampler_valid:
batch_ordering: !ref <batch_ordering>
max_batch_ex: !ref <max_batch_ex>
+
# Dataloader options
train_dataloader_opts:
batch_size: !ref <batch_size>
@@ -99,7 +101,8 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-####################### Model parameters ###########################
+####################### Model Parameters #######################################
+
# Transformer
d_model: 256
nhead: 8
@@ -132,7 +135,7 @@ test_beam_size: 66
lm_weight: 0.60
ctc_weight_decode: 0.40
-############################## models ################################
+############################## Models ##########################################
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
input_shape: (8, 10, 80)
@@ -203,7 +206,8 @@ Adam: !name:torch.optim.Adam
betas: (0.9, 0.98)
eps: 0.000000001
-# Scorer
+####################### Decoding & optimiser ###################################
+
ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
eos_index: !ref <eos_index>
blank_index: !ref <blank_index>
@@ -273,57 +277,34 @@ checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 4 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 4 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "mean" # Method of dropping chunks
-
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
-
-# Frequency Drop
-freq_drop_length_low: 10 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 20 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 4 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 4 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "mean" # Method of dropping chunks
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
+# Freq Drop
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 10
+ drop_length_high: 20
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -337,6 +318,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Logging and Pretrainer ##########################
+
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
save_file: !ref <train_log>
diff --git a/recipes/LibriSpeech/ASR/transformer/hparams/hyperconformer_8M.yaml b/recipes/LibriSpeech/ASR/transformer/hparams/hyperconformer_8M.yaml
index 47053b5ff..fe3bd599c 100644
--- a/recipes/LibriSpeech/ASR/transformer/hparams/hyperconformer_8M.yaml
+++ b/recipes/LibriSpeech/ASR/transformer/hparams/hyperconformer_8M.yaml
@@ -40,7 +40,8 @@ test_csv:
- !ref <output_folder>/test-clean.csv
- !ref <output_folder>/test-other.csv
-# Training parameters
+####################### Training Parameters ####################################
+
# To make Transformers converge, the global bath size should be large enough.
# The global batch size is computed as batch_size * n_gpus * grad_accumulation_factor.
# Empirically, we found that this value should be >= 128.
@@ -100,7 +101,8 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-####################### Model parameters ###########################
+####################### Model Parameters #######################################
+
# Transformer
d_model: 144
nhead: 8
@@ -133,7 +135,7 @@ test_beam_size: 66
lm_weight: 0.60
ctc_weight_decode: 0.40
-############################## models ################################
+############################## Models ##########################################
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
input_shape: (8, 10, 80)
@@ -204,7 +206,8 @@ Adam: !name:torch.optim.Adam
betas: (0.9, 0.98)
eps: 0.000000001
-# Scorer
+####################### Decoding & optimiser ###########################
+
ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
eos_index: !ref <eos_index>
blank_index: !ref <blank_index>
@@ -274,57 +277,34 @@ checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 4 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 4 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "mean" # Method of dropping chunks
-
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
-
-# Frequency Drop
-freq_drop_length_low: 10 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 20 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 4 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 4 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "mean" # Method of dropping chunks
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
+# Freq Drop
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 10
+ drop_length_high: 20
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -338,6 +318,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Logging and Pretrainer ##########################
+
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
save_file: !ref <train_log>
diff --git a/recipes/LibriSpeech/ASR/transformer/hparams/train_hf_whisper.yaml b/recipes/LibriSpeech/ASR/transformer/hparams/train_hf_whisper.yaml
index 4805d7c6c..4891ca617 100644
--- a/recipes/LibriSpeech/ASR/transformer/hparams/train_hf_whisper.yaml
+++ b/recipes/LibriSpeech/ASR/transformer/hparams/train_hf_whisper.yaml
@@ -35,7 +35,8 @@ test_csv:
ckpt_interval_minutes: 30 # save checkpoint every N min
-# Training parameters
+############################## Training Parameters #############################
+
number_of_epochs: 1
lr_whisper: 0.00003
sorting: ascending
@@ -61,7 +62,7 @@ min_decode_ratio: 0.0
max_decode_ratio: 1.0
test_beam_size: 8
-# Model parameters
+####################### Model Parameters #######################################
freeze_whisper: False
@@ -74,52 +75,34 @@ valid_loader_kwargs:
test_loader_kwargs:
batch_size: !ref <test_batch_size>
-
-#
-# Functions and classes
-#
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0 # Min frequency band dropout probability
+ drop_freq_high: 1 # Max frequency band dropout probability
+ drop_freq_count_low: 1 # Min number of frequency bands to drop
+ drop_freq_count_high: 3 # Max number of frequency bands to drop
+ drop_freq_width: 0.05 # Width of frequency bands to drop
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1
+ drop_length_high: 5
+ drop_count_low: 1000
+ drop_count_high: 2000
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -128,6 +111,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
whisper: !new:speechbrain.lobes.models.huggingface_transformers.whisper.Whisper
source: !ref <whisper_hub>
freeze: !ref <freeze_whisper>
@@ -142,6 +127,8 @@ nll_loss: !name:speechbrain.nnet.losses.nll_loss
modules:
whisper: !ref <whisper>
+############################## Decoding & optimiser ############################
+
whisper_opt_class: !name:torch.optim.AdamW
lr: !ref <lr_whisper>
weight_decay: 0.01
@@ -167,6 +154,8 @@ lr_annealing_whisper: !new:speechbrain.nnet.schedulers.NewBobScheduler
annealing_factor: 0.9
patient: 0
+############################## Logging and Pretrainer ##########################
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
checkpoints_dir: !ref <save_folder>
recoverables:
diff --git a/recipes/LibriSpeech/ASR/transformer/hparams/transformer.yaml b/recipes/LibriSpeech/ASR/transformer/hparams/transformer.yaml
index 36626f0d9..173453e9d 100644
--- a/recipes/LibriSpeech/ASR/transformer/hparams/transformer.yaml
+++ b/recipes/LibriSpeech/ASR/transformer/hparams/transformer.yaml
@@ -42,7 +42,8 @@ test_csv:
ckpt_interval_minutes: 30 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
+
# To make Transformers converge, the global bath size should be large enough.
# The global batch size is computed as batch_size * n_gpus * grad_accumulation_factor.
# Empirically, we found that this value should be >= 128.
@@ -116,7 +117,8 @@ test_dataloader_opts:
padding_kwargs:
value: !ref <pad_index>
-####################### Model parameters ###########################
+####################### Model Parameters #######################################
+
# Transformer
d_model: 512
nhead: 4
@@ -142,7 +144,7 @@ test_beam_size: 66
lm_weight: 0.60
ctc_weight_decode: 0.40
-############################## models ################################
+############################## Models ##########################################
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
input_shape: (8, 10, 80)
@@ -210,7 +212,8 @@ Adam: !name:torch.optim.Adam
eps: 0.000000001
-# Scorer
+####################### Decoding & optimiser ###################################
+
ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
eos_index: !ref <eos_index>
blank_index: !ref <blank_index>
@@ -284,57 +287,34 @@ normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
update_until_epoch: 4
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 4 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 4 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "mean" # Method of dropping chunks
-
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
-
-# Frequency Drop
-freq_drop_length_low: 10 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 20 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 4 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 4 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "mean" # Method of dropping chunks
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
+# Freq Drop
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 10
+ drop_length_high: 20
+ drop_count_low: 4
+ drop_count_high: 4
+ replace: "mean"
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -348,6 +328,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Logging and Pretrainer ##########################
+
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
save_file: !ref <train_log>
diff --git a/recipes/LibriSpeech/ASR/transformer/train.py b/recipes/LibriSpeech/ASR/transformer/train.py
index b69763e26..292d7cc42 100644
--- a/recipes/LibriSpeech/ASR/transformer/train.py
+++ b/recipes/LibriSpeech/ASR/transformer/train.py
@@ -114,16 +114,16 @@ class ASR(sb.core.Brain):
tokens, tokens_lens = batch.tokens
if stage == sb.Stage.TRAIN:
+ # Labels must be extended if parallel augmentation or concatenated
+ # augmentation was performed on the input (increasing the time dimension)
if hasattr(self.hparams, "fea_augment"):
- tokens = self.hparams.fea_augment.replicate_labels(tokens)
- tokens_lens = self.hparams.fea_augment.replicate_labels(
- tokens_lens
- )
- tokens_eos = self.hparams.fea_augment.replicate_labels(
- tokens_eos
- )
- tokens_eos_lens = self.hparams.fea_augment.replicate_labels(
- tokens_eos_lens
+ (
+ tokens,
+ tokens_lens,
+ tokens_eos,
+ tokens_eos_lens,
+ ) = self.hparams.fea_augment.replicate_multiple_labels(
+ tokens, tokens_lens, tokens_eos, tokens_eos_lens
)
loss_seq = self.hparams.seq_cost(
diff --git a/recipes/LibriSpeech/G2P/hparams/hparams_g2p_rnn.yaml b/recipes/LibriSpeech/G2P/hparams/hparams_g2p_rnn.yaml
index a565905c8..f487ffbe1 100644
--- a/recipes/LibriSpeech/G2P/hparams/hparams_g2p_rnn.yaml
+++ b/recipes/LibriSpeech/G2P/hparams/hparams_g2p_rnn.yaml
@@ -95,7 +95,7 @@ homograph_loss_weight: 2.0
lr: 0.002
save_for_pretrained: True
-# Model parameters
+####################### Model Parameters #######################################
output_neurons: !apply:speechbrain.utils.hparams.choice
value: !ref <phn_tokenize>
choices:
diff --git a/recipes/LibriSpeech/G2P/hparams/hparams_g2p_transformer.yaml b/recipes/LibriSpeech/G2P/hparams/hparams_g2p_transformer.yaml
index c75cd97bf..e1c0f44c7 100644
--- a/recipes/LibriSpeech/G2P/hparams/hparams_g2p_transformer.yaml
+++ b/recipes/LibriSpeech/G2P/hparams/hparams_g2p_transformer.yaml
@@ -95,7 +95,7 @@ lr_dont_halve_until_epoch: 1
lr_patience: 1
save_for_pretrained: True
-# Model parameters
+####################### Model Parameters #######################################
output_neurons: !apply:speechbrain.utils.hparams.choice
value: !ref <phn_tokenize>
choices:
diff --git a/recipes/LibriSpeech/G2P/hparams/hparams_lm_rnn.yaml b/recipes/LibriSpeech/G2P/hparams/hparams_lm_rnn.yaml
index dcb768259..7e1b7bc4a 100644
--- a/recipes/LibriSpeech/G2P/hparams/hparams_lm_rnn.yaml
+++ b/recipes/LibriSpeech/G2P/hparams/hparams_lm_rnn.yaml
@@ -50,7 +50,7 @@ train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
tokenizer_file: <output_folder>/save/phoneme_tokenizer.model
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 20
batch_size: 80
lr: 0.001
@@ -68,7 +68,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-# Model parameters
+####################### Model Parameters #######################################
model_dim: !apply:speechbrain.utils.hparams.choice
value: !ref <phn_tokenize>
choices:
diff --git a/recipes/LibriSpeech/G2P/hparams/hparams_lm_transformer.yaml b/recipes/LibriSpeech/G2P/hparams/hparams_lm_transformer.yaml
index 2a9a434d6..5e319e3d8 100644
--- a/recipes/LibriSpeech/G2P/hparams/hparams_lm_transformer.yaml
+++ b/recipes/LibriSpeech/G2P/hparams/hparams_lm_transformer.yaml
@@ -39,7 +39,7 @@ train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
# Tokenizer model (you must use the same tokenizer for LM and ASR training)
tokenizer_file: <output_folder>/save/phoneme_tokenizer.model
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 20
batch_size: 80
lr: 0.001
@@ -57,7 +57,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-# Model parameters
+####################### Model Parameters #######################################
emb_dim: 256 # dimension of the embeddings
transformer_num_heads: 4
diff --git a/recipes/LibriSpeech/LM/hparams/RNNLM.yaml b/recipes/LibriSpeech/LM/hparams/RNNLM.yaml
index b061b4fcc..0896de960 100644
--- a/recipes/LibriSpeech/LM/hparams/RNNLM.yaml
+++ b/recipes/LibriSpeech/LM/hparams/RNNLM.yaml
@@ -29,7 +29,7 @@ test_transcripts_pattern: "test*/**/*.trans.txt"
# Tokenizer model
tokenizer_file: https://www.dropbox.com/s/o7gnouwdoqchotj/1000_unigram.model?dl=1
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 20
batch_size: 80
lr: 0.001
@@ -47,7 +47,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-# Model parameters
+####################### Model Parameters #######################################
emb_size: 128
activation: !name:torch.nn.LeakyReLU
dropout: 0.0
diff --git a/recipes/LibriSpeech/LM/hparams/transformer.yaml b/recipes/LibriSpeech/LM/hparams/transformer.yaml
index c79ef5769..50123a4c3 100644
--- a/recipes/LibriSpeech/LM/hparams/transformer.yaml
+++ b/recipes/LibriSpeech/LM/hparams/transformer.yaml
@@ -29,7 +29,7 @@ test_transcripts_pattern: "test*/**/*.trans.txt"
# Tokenizer model
tokenizer_file: speechbrain/asr-transformer-transformerlm-librispeech/tokenizer.ckpt
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 20
batch_size: 16
lr: 10
diff --git a/recipes/LibriSpeech/Tokenizer/hparams/1K_unigram_subword_bpe.yaml b/recipes/LibriSpeech/Tokenizer/hparams/1K_unigram_subword_bpe.yaml
index b5a9fa60e..9dda21f82 100644
--- a/recipes/LibriSpeech/Tokenizer/hparams/1K_unigram_subword_bpe.yaml
+++ b/recipes/LibriSpeech/Tokenizer/hparams/1K_unigram_subword_bpe.yaml
@@ -16,7 +16,7 @@ skip_prep: False
train_csv: !ref <output_folder>/train.csv
valid_csv: !ref <output_folder>/dev-clean.csv
-# Training parameters
+####################### Training Parameters ####################################
token_type: unigram # ["unigram", "bpe", "char"]
token_output: 1000 # index(blank/eos/bos/unk) = 0
character_coverage: 1.0
diff --git a/recipes/LibriSpeech/Tokenizer/hparams/5K_unigram_subword_bpe.yaml b/recipes/LibriSpeech/Tokenizer/hparams/5K_unigram_subword_bpe.yaml
index c312ce5bb..1f328c6f1 100644
--- a/recipes/LibriSpeech/Tokenizer/hparams/5K_unigram_subword_bpe.yaml
+++ b/recipes/LibriSpeech/Tokenizer/hparams/5K_unigram_subword_bpe.yaml
@@ -16,7 +16,7 @@ skip_prep: False
train_csv: !ref <output_folder>/train.csv
valid_csv: !ref <output_folder>/dev-clean.csv
-# Training parameters
+####################### Training Parameters ####################################
token_type: unigram # ["unigram", "bpe", "char"]
token_output: 5000 # index(blank/eos/bos/unk) = 0
character_coverage: 1.0
diff --git a/recipes/LibriSpeech/self-supervised-learning/wav2vec2/hparams/wav2vec2_base.yaml b/recipes/LibriSpeech/self-supervised-learning/wav2vec2/hparams/wav2vec2_base.yaml
index 4806f3e06..13ce0d220 100644
--- a/recipes/LibriSpeech/self-supervised-learning/wav2vec2/hparams/wav2vec2_base.yaml
+++ b/recipes/LibriSpeech/self-supervised-learning/wav2vec2/hparams/wav2vec2_base.yaml
@@ -48,7 +48,7 @@ test_dataloader_options:
batch_size: 8 # DynamicBatching not used at testing time
num_workers: 4
-# Training parameters
+####################### Training Parameters ####################################
lr: 0.0005
warmup: 30000
# This is equivalent to optimizer_step_limit - warmup
@@ -63,7 +63,7 @@ mask_prob: 0.65
mask_length: 10
num_negatives: 100
-# Model parameters
+####################### Model Parameters #######################################
embedding_dim: 768
extractor_dim: 512
final_dim: 256
diff --git a/recipes/MEDIA/ASR/CTC/hparams/train_hf_wav2vec.yaml b/recipes/MEDIA/ASR/CTC/hparams/train_hf_wav2vec.yaml
index 924579a9c..70ef38de7 100644
--- a/recipes/MEDIA/ASR/CTC/hparams/train_hf_wav2vec.yaml
+++ b/recipes/MEDIA/ASR/CTC/hparams/train_hf_wav2vec.yaml
@@ -55,7 +55,7 @@ test_dataloader_options:
sample_rate: 16000
feats_dim: 1024
-# Training parameters:
+####################### Training Parameters ####################################:
number_of_epochs: 30
lr: 1
lr_wav2vec: 0.0001
@@ -67,7 +67,7 @@ patient: 0
patient_wav2vec: 0
sorting: ascending
-# Model parameters:
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dnn_blocks: 3
dnn_neurons: 512
diff --git a/recipes/MEDIA/SLU/CTC/hparams/train_hf_wav2vec_full.yaml b/recipes/MEDIA/SLU/CTC/hparams/train_hf_wav2vec_full.yaml
index 7be6f6b94..4f9bad2e7 100644
--- a/recipes/MEDIA/SLU/CTC/hparams/train_hf_wav2vec_full.yaml
+++ b/recipes/MEDIA/SLU/CTC/hparams/train_hf_wav2vec_full.yaml
@@ -57,7 +57,7 @@ test_dataloader_options:
sample_rate: 16000
feats_dim: 1024
-# Training parameters:
+####################### Training Parameters ####################################:
number_of_epochs: 30
lr: 1
lr_wav2vec: 0.0001
@@ -69,7 +69,7 @@ patient: 0
patient_wav2vec: 0
sorting: ascending
-# Model parameters:
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dnn_blocks: 3
dnn_neurons: 512
diff --git a/recipes/MEDIA/SLU/CTC/hparams/train_hf_wav2vec_relax.yaml b/recipes/MEDIA/SLU/CTC/hparams/train_hf_wav2vec_relax.yaml
index d631a6da8..8631e6e88 100644
--- a/recipes/MEDIA/SLU/CTC/hparams/train_hf_wav2vec_relax.yaml
+++ b/recipes/MEDIA/SLU/CTC/hparams/train_hf_wav2vec_relax.yaml
@@ -57,7 +57,7 @@ test_dataloader_options:
sample_rate: 16000
feats_dim: 1024
-# Training parameters:
+####################### Training Parameters ####################################:
number_of_epochs: 30
lr: 1
lr_wav2vec: 0.0001
@@ -69,7 +69,7 @@ patient: 0
patient_wav2vec: 0
sorting: ascending
-# Model parameters:
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dnn_blocks: 3
dnn_neurons: 512
diff --git a/recipes/MultiWOZ/response_generation/gpt/hparams/train_gpt.yaml b/recipes/MultiWOZ/response_generation/gpt/hparams/train_gpt.yaml
index 05b16bea3..5bb3b8ed8 100644
--- a/recipes/MultiWOZ/response_generation/gpt/hparams/train_gpt.yaml
+++ b/recipes/MultiWOZ/response_generation/gpt/hparams/train_gpt.yaml
@@ -58,7 +58,7 @@ max_history: 5
ignore_index: -100
label_smoothing: 0
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 4
batch_size: 8
test_batch_size: 4
diff --git a/recipes/MultiWOZ/response_generation/llama2/hparams/train_llama2.yaml b/recipes/MultiWOZ/response_generation/llama2/hparams/train_llama2.yaml
index f7fd2b087..507115e83 100644
--- a/recipes/MultiWOZ/response_generation/llama2/hparams/train_llama2.yaml
+++ b/recipes/MultiWOZ/response_generation/llama2/hparams/train_llama2.yaml
@@ -40,7 +40,7 @@ max_history: 2
ignore_index: -100
label_smoothing: 0
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 4
batch_size: 1
test_batch_size: 1
diff --git a/recipes/REAL-M/sisnr-estimation/hparams/pool_sisnrestimator.yaml b/recipes/REAL-M/sisnr-estimation/hparams/pool_sisnrestimator.yaml
index 3f1da0919..c23c11c53 100644
--- a/recipes/REAL-M/sisnr-estimation/hparams/pool_sisnrestimator.yaml
+++ b/recipes/REAL-M/sisnr-estimation/hparams/pool_sisnrestimator.yaml
@@ -67,7 +67,7 @@ num_spks: 2
noprogressbar: False
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.0001
diff --git a/recipes/RescueSpeech/ASR/noise-robust/hparams/robust_asr_16k.yaml b/recipes/RescueSpeech/ASR/noise-robust/hparams/robust_asr_16k.yaml
index 302dabe57..10e8e58e4 100644
--- a/recipes/RescueSpeech/ASR/noise-robust/hparams/robust_asr_16k.yaml
+++ b/recipes/RescueSpeech/ASR/noise-robust/hparams/robust_asr_16k.yaml
@@ -41,7 +41,7 @@ skip_prep: False
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
-## Model parameters- Enhance model
+## Model Parameters- Enhance model
dereverberate: False
save_audio: True
sample_rate: 16000
@@ -54,7 +54,7 @@ use_rand_shift: False
min_shift: -8000
max_shift: 8000
-## Training parameters- ASR
+######################## Training Parameters ####################################- ASR
number_of_epochs: 10
lr_whisper: 0.00003
sorting: ascending
diff --git a/recipes/SLURP/NLU/hparams/train.yaml b/recipes/SLURP/NLU/hparams/train.yaml
index e2201d96b..7d88d62a9 100644
--- a/recipes/SLURP/NLU/hparams/train.yaml
+++ b/recipes/SLURP/NLU/hparams/train.yaml
@@ -28,14 +28,14 @@ asr_tokenizer_file: https://www.dropbox.com/s/o7gnouwdoqchotj/1000_unigram.model
slu_tokenizer_file: https://www.dropbox.com/s/tmwq12r5vgcsif9/58_unigram.model?dl=1
skip_prep: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 20
batch_size: 16
lr: 0.0003
# token_type: unigram # ["unigram", "bpe", "char"]
sorting: random
-# Model parameters
+####################### Model Parameters #######################################
# sample_rate: 1600
emb_size: 128
dec_neurons: 512
diff --git a/recipes/SLURP/Tokenizer/hparams/tokenizer_bpe58.yaml b/recipes/SLURP/Tokenizer/hparams/tokenizer_bpe58.yaml
index 51f805b07..bf935024a 100644
--- a/recipes/SLURP/Tokenizer/hparams/tokenizer_bpe58.yaml
+++ b/recipes/SLURP/Tokenizer/hparams/tokenizer_bpe58.yaml
@@ -14,7 +14,7 @@ train_csv: !ref <output_folder>/train-type=direct.csv
valid_csv: !ref <output_folder>/devel-type=direct.csv
skip_prep: False
-# Training parameters
+####################### Training Parameters ####################################
token_type: unigram # ["unigram", "bpe", "char"]
token_output: 58 # index(blank/eos/bos/unk) = 0
character_coverage: 1.0
diff --git a/recipes/SLURP/direct/hparams/train.yaml b/recipes/SLURP/direct/hparams/train.yaml
index 5a42c738c..038d2e59e 100644
--- a/recipes/SLURP/direct/hparams/train.yaml
+++ b/recipes/SLURP/direct/hparams/train.yaml
@@ -34,7 +34,7 @@ rir_annotation: !ref <save_folder>/rir.csv
tokenizer_file: https://www.dropbox.com/s/tmwq12r5vgcsif9/58_unigram.model?dl=1
skip_prep: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 20
batch_size: 16
lr: 0.0003
@@ -42,7 +42,7 @@ lr: 0.0003
sorting: random
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Model parameters
+####################### Model Parameters #######################################
sample_rate: 16000
emb_size: 128
dec_neurons: 512
@@ -100,41 +100,31 @@ add_noise: !new:speechbrain.augment.time_domain.AddNoise
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
+############################## Augmentations ###################################
+
# Speed perturbation
-speed_changes: [90, 95, 105, 110] # List of speed changes for time-stretching
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [90, 95, 105, 110]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 3 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 3
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
shuffle_augmentations: True
min_augmentations: 1
max_augmentations: 4
@@ -146,7 +136,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
-# Models
+############################## Models ##########################################
+
asr_model_source: speechbrain/asr-crdnn-rnnlm-librispeech
slu_enc: !new:speechbrain.nnet.containers.Sequential
diff --git a/recipes/SLURP/direct/hparams/train_with_wav2vec2.yaml b/recipes/SLURP/direct/hparams/train_with_wav2vec2.yaml
index b383da5cb..84222db5f 100644
--- a/recipes/SLURP/direct/hparams/train_with_wav2vec2.yaml
+++ b/recipes/SLURP/direct/hparams/train_with_wav2vec2.yaml
@@ -32,7 +32,7 @@ skip_prep: False
# URL for the wav2vec2 model, you can change to benchmark diffrenet models
wav2vec2_hub: "facebook/hubert-base-ls960"
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 35
batch_size: 6
lr: 0.0003
@@ -47,7 +47,7 @@ freeze_wav2vec2: False
#set to true to freeze the CONV part of the wav2vec2 model
freeze_wav2vec2_conv: True
-# Model parameters
+####################### Model Parameters #######################################
sample_rate: 16000
emb_size: 128
dec_neurons: 512
@@ -96,45 +96,31 @@ seq_lin: !new:speechbrain.nnet.linear.Linear
input_size: !ref <dec_neurons>
n_neurons: !ref <output_neurons>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
diff --git a/recipes/Switchboard/ASR/CTC/hparams/train_with_wav2vec.yaml b/recipes/Switchboard/ASR/CTC/hparams/train_with_wav2vec.yaml
index 2933e0fd4..7741680bd 100644
--- a/recipes/Switchboard/ASR/CTC/hparams/train_with_wav2vec.yaml
+++ b/recipes/Switchboard/ASR/CTC/hparams/train_with_wav2vec.yaml
@@ -49,7 +49,7 @@ test_csv:
- !ref <output_folder>/test_callhome.csv
- !ref <output_folder>/test.csv
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
lr: 1.0
lr_wav2vec: 0.0001
@@ -74,7 +74,7 @@ test_dataloader_options:
token_type: unigram # ["unigram", "bpe", "char"]
character_coverage: 1.0
-# Model parameters
+####################### Model Parameters #######################################
wav2vec_output_dim: 1024
dnn_neurons: 1024
freeze_wav2vec: False
@@ -109,45 +109,31 @@ kenlm_model_path: null
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -156,6 +142,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.nnet.containers.Sequential
input_shape: [null, null, !ref <wav2vec_output_dim>]
linear1: !name:speechbrain.nnet.linear.Linear
diff --git a/recipes/Switchboard/ASR/seq2seq/hparams/train_BPE_2000.yaml b/recipes/Switchboard/ASR/seq2seq/hparams/train_BPE_2000.yaml
index d20001f1c..743467bcf 100644
--- a/recipes/Switchboard/ASR/seq2seq/hparams/train_BPE_2000.yaml
+++ b/recipes/Switchboard/ASR/seq2seq/hparams/train_BPE_2000.yaml
@@ -57,7 +57,7 @@ test_csv:
- !ref <save_folder>/test_callhome.csv
- !ref <save_folder>/test.csv
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 20
number_of_ctc_epochs: 5
batch_size: 10
@@ -103,7 +103,7 @@ test_dataloader_opts:
num_workers: !ref <num_workers>
batch_size: !ref <batch_size>
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 2
@@ -144,57 +144,40 @@ prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_U
ext: wav
csv_file: !ref <noise_annotation>
-# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
+############################## Augmentations ###################################
+# Add noise to input signal
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -215,6 +198,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
+############################## Models ##########################################
+
enc: !new:speechbrain.lobes.models.CRDNN.CRDNN
input_shape: [null, null, !ref <n_mels>]
activation: !ref <activation>
diff --git a/recipes/Switchboard/ASR/seq2seq/train.py b/recipes/Switchboard/ASR/seq2seq/train.py
index 57ce5c966..d0cd3ce91 100644
--- a/recipes/Switchboard/ASR/seq2seq/train.py
+++ b/recipes/Switchboard/ASR/seq2seq/train.py
@@ -127,12 +127,16 @@ class ASR(sb.Brain):
tokens_eos, tokens_eos_lens = batch.tokens_eos
tokens, tokens_lens = batch.tokens
+ # Labels must be extended if parallel augmentation or concatenated
+ # augmentation was performed on the input (increasing the time dimension)
if stage == sb.Stage.TRAIN and hasattr(self.hparams, "wav_augment"):
- tokens = self.hparams.wav_augment.replicate_labels(tokens)
- tokens_lens = self.hparams.wav_augment.replicate_labels(tokens_lens)
- tokens_eos = self.hparams.wav_augment.replicate_labels(tokens_eos)
- tokens_eos_lens = self.hparams.wav_augment.replicate_labels(
- tokens_eos_lens
+ (
+ tokens,
+ tokens_lens,
+ tokens_eos,
+ tokens_eos_lens,
+ ) = self.hparams.wav_augment.replicate_multiple_labels(
+ tokens, tokens_lens, tokens_eos, tokens_eos_lens
)
loss_seq = self.hparams.seq_cost(
diff --git a/recipes/Switchboard/ASR/transformer/hparams/transformer.yaml b/recipes/Switchboard/ASR/transformer/hparams/transformer.yaml
index bd84a8a19..674c03719 100644
--- a/recipes/Switchboard/ASR/transformer/hparams/transformer.yaml
+++ b/recipes/Switchboard/ASR/transformer/hparams/transformer.yaml
@@ -51,7 +51,7 @@ test_csv:
ckpt_interval_minutes: 30 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
# To make Transformers converge, the global batch size should be large enough.
# The global batch size is computed as:
# batch_size * n_gpus * grad_accumulation_factor.
@@ -96,7 +96,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-####################### Model parameters ###########################
+####################### Model Parameters ###########################
# Transformer
transformer_input_size: 1280
d_model: 256
@@ -271,50 +271,32 @@ normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
update_until_epoch: 4
-# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
+############################## Augmentations ###################################
+# Speed perturbation
+speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
+ orig_freq: !ref <sample_rate>
+ speeds: [95, 100, 105]
+
+# Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 2
+ drop_count_high: 2
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -323,14 +305,7 @@ fea_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <freq_drop>,
!ref <time_warp>]
-
-# Speed perturbation
do_speed_perturb: True
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
-speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
- orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
compute_features: !new:speechbrain.lobes.features.Fbank
sample_rate: !ref <sample_rate>
diff --git a/recipes/Switchboard/ASR/transformer/hparams/transformer_finetuned_LM.yaml b/recipes/Switchboard/ASR/transformer/hparams/transformer_finetuned_LM.yaml
index 45a765cd1..8dd221ca4 100644
--- a/recipes/Switchboard/ASR/transformer/hparams/transformer_finetuned_LM.yaml
+++ b/recipes/Switchboard/ASR/transformer/hparams/transformer_finetuned_LM.yaml
@@ -51,7 +51,7 @@ test_csv:
ckpt_interval_minutes: 30 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
# To make Transformers converge, the global batch size should be large enough.
# The global batch size is computed as:
# batch_size * n_gpus * grad_accumulation_factor.
@@ -96,7 +96,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-####################### Model parameters ###########################
+####################### Model Parameters ###########################
# Transformer
d_model: 512
nhead: 4
@@ -126,7 +126,7 @@ lm_weight: 0.60
ctc_weight_decode: 0.40
temperature: 1.15
temperature_lm: 1.15
-############################## models ################################
+############################## Models ################################
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
input_shape: (8, 10, 80)
@@ -258,57 +258,32 @@ normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
update_until_epoch: 4
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Time Drop
-time_drop_length_low: 15 # Min length for temporal chunk to drop in spectrogram
-time_drop_length_high: 25 # Max length for temporal chunk to drop in spectrogram
-time_drop_count_low: 5 # Min number of chunks to drop in time in the spectrogram
-time_drop_count_high: 5 # Max number of chunks to drop in time in the spectrogram
-time_drop_replace: "zeros" # Method of dropping chunks
-
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <time_drop_length_low>
- drop_length_high: !ref <time_drop_length_high>
- drop_count_low: !ref <time_drop_count_low>
- drop_count_high: !ref <time_drop_count_high>
- replace: !ref <time_drop_replace>
- dim: 1
+ drop_length_low: 15
+ drop_length_high: 25
+ drop_count_low: 5
+ drop_count_high: 5
# Frequency Drop
-freq_drop_length_low: 25 # Min length for chunks to drop in frequency in the spectrogram
-freq_drop_length_high: 35 # Max length for chunks to drop in frequency in the spectrogram
-freq_drop_count_low: 2 # Min number of chunks to drop in frequency in the spectrogram
-freq_drop_count_high: 2 # Max number of chunks to drop in frequency in the spectrogram
-freq_drop_replace: "zeros" # Method of dropping chunks
-
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
- drop_length_low: !ref <freq_drop_length_low>
- drop_length_high: !ref <freq_drop_length_high>
- drop_count_low: !ref <freq_drop_count_low>
- drop_count_high: !ref <freq_drop_count_high>
- replace: !ref <freq_drop_replace>
+ drop_length_low: 25
+ drop_length_high: 35
+ drop_count_low: 2
+ drop_count_high: 2
dim: 2
# Time warp
-time_warp_window: 5 # Length of time warping window
-time_warp_mode: "bicubic" # Time warping method
-
time_warp: !new:speechbrain.augment.freq_domain.Warping
- warp_window: !ref <time_warp_window>
- warp_mode: !ref <time_warp_mode>
- dim: 1
fea_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
diff --git a/recipes/Switchboard/ASR/transformer/train.py b/recipes/Switchboard/ASR/transformer/train.py
index 5fb6ebc47..dcc527952 100644
--- a/recipes/Switchboard/ASR/transformer/train.py
+++ b/recipes/Switchboard/ASR/transformer/train.py
@@ -133,16 +133,16 @@ class ASR(sb.core.Brain):
tokens, tokens_lens = batch.tokens
if stage == sb.Stage.TRAIN:
+ # Labels must be extended if parallel augmentation or concatenated
+ # augmentation was performed on the input (increasing the time dimension)
if hasattr(self.hparams, "fea_augment"):
- tokens = self.hparams.fea_augment.replicate_labels(tokens)
- tokens_lens = self.hparams.fea_augment.replicate_labels(
- tokens_lens
- )
- tokens_eos = self.hparams.fea_augment.replicate_labels(
- tokens_eos
- )
- tokens_eos_lens = self.hparams.fea_augment.replicate_labels(
- tokens_eos_lens
+ (
+ tokens,
+ tokens_lens,
+ tokens_eos,
+ tokens_eos_lens,
+ ) = self.hparams.fea_augment.replicate_multiple_labels(
+ tokens, tokens_lens, tokens_eos, tokens_eos_lens
)
loss_seq = self.hparams.seq_cost(
diff --git a/recipes/Switchboard/LM/hparams/transformer.yaml b/recipes/Switchboard/LM/hparams/transformer.yaml
index 2f27463af..b501faf55 100644
--- a/recipes/Switchboard/LM/hparams/transformer.yaml
+++ b/recipes/Switchboard/LM/hparams/transformer.yaml
@@ -36,7 +36,7 @@ test_csv: !ref <save_folder>/test.csv
# (e.g. /path/to/2000_unigram.model)
tokenizer_file: !PLACEHOLDER
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 100
batch_size: 164
lr: 1
diff --git a/recipes/Switchboard/LM/hparams/transformer_finetune.yaml b/recipes/Switchboard/LM/hparams/transformer_finetune.yaml
index f5657c76c..5b0860e41 100644
--- a/recipes/Switchboard/LM/hparams/transformer_finetune.yaml
+++ b/recipes/Switchboard/LM/hparams/transformer_finetune.yaml
@@ -39,7 +39,7 @@ test_csv: !ref <save_folder>/test.csv
# instead. E.g if you want to use your own LM / tokenizer.
pretrained_lm_tokenizer_path: speechbrain/asr-transformer-transformerlm-librispeech
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 5
batch_size: 128
lr: 2
diff --git a/recipes/Switchboard/Tokenizer/hparams/2K_unigram_subword_bpe.yaml b/recipes/Switchboard/Tokenizer/hparams/2K_unigram_subword_bpe.yaml
index e6c546bdf..d07d83e70 100644
--- a/recipes/Switchboard/Tokenizer/hparams/2K_unigram_subword_bpe.yaml
+++ b/recipes/Switchboard/Tokenizer/hparams/2K_unigram_subword_bpe.yaml
@@ -20,7 +20,7 @@ train_csv: !ref <output_folder>/train_lm.csv
valid_csv: !ref <output_folder>/dev.csv
skip_prep: False
-# Training parameters
+####################### Training Parameters ####################################
token_type: unigram # ["unigram", "bpe", "char"]
token_output: 2000 # index(blank/eos/bos/unk) = 0
character_coverage: 1.0
diff --git a/recipes/TIMIT/ASR/CTC/hparams/train.yaml b/recipes/TIMIT/ASR/CTC/hparams/train.yaml
index dce350b7e..145fa1a3e 100644
--- a/recipes/TIMIT/ASR/CTC/hparams/train.yaml
+++ b/recipes/TIMIT/ASR/CTC/hparams/train.yaml
@@ -25,7 +25,7 @@ data_folder_noise: !ref <data_folder>/noise # The noisy sequencies for data augm
NOISE_DATASET_URL: https://www.dropbox.com/scl/fi/a09pj97s5ifan81dqhi4n/noises.zip?rlkey=j8b0n9kdjdr32o1f06t0cw5b7&dl=1
noise_annotation: !ref <save_folder>/noise.csv #The data manifest files are created by the data preparation script
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 50
batch_size: 8
lr: 1.0
@@ -36,7 +36,7 @@ sample_rate: 16000
n_fft: 400
n_mels: 40
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 2
@@ -69,6 +69,8 @@ test_dataloader_opts:
normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
+############################## Augmentations ###################################
+
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
URL: !ref <NOISE_DATASET_URL>
@@ -76,58 +78,38 @@ prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_U
ext: wav
csv_file: !ref <noise_annotation>
-
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -137,6 +119,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
diff --git a/recipes/TIMIT/ASR/seq2seq/hparams/train.yaml b/recipes/TIMIT/ASR/seq2seq/hparams/train.yaml
index cc3276506..d61179fa9 100644
--- a/recipes/TIMIT/ASR/seq2seq/hparams/train.yaml
+++ b/recipes/TIMIT/ASR/seq2seq/hparams/train.yaml
@@ -22,7 +22,7 @@ test_annotation: !ref <save_folder>/test.json
skip_prep: False # Skip data preparation
uppercase: False # Must be True when the TIMIT dataset is in the upper-case version
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 50
batch_size: 8 # Used if dynamic_batching is False
lr: 0.0003
@@ -34,7 +34,7 @@ sample_rate: 16000
n_fft: 400
n_mels: 40
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 2
@@ -88,45 +88,30 @@ dynamic_batch_sampler:
shuffle: !ref <shuffle>
batch_ordering: !ref <batch_ordering>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -135,6 +120,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
diff --git a/recipes/TIMIT/ASR/seq2seq/hparams/train_with_wav2vec2.yaml b/recipes/TIMIT/ASR/seq2seq/hparams/train_with_wav2vec2.yaml
index 9756e9356..705f79e9a 100644
--- a/recipes/TIMIT/ASR/seq2seq/hparams/train_with_wav2vec2.yaml
+++ b/recipes/TIMIT/ASR/seq2seq/hparams/train_with_wav2vec2.yaml
@@ -23,7 +23,7 @@ test_annotation: !ref <save_folder>/test.json
skip_prep: False # Skip data preparation
uppercase: False # Must be True when the TIMIT dataset is in the upper-case version
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 20
batch_size: 8
lr: 0.0003
@@ -33,7 +33,7 @@ sorting: ascending
precision: fp32 # bf16, fp16 or fp32
sample_rate: 16000
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dnn_layers: 2
dnn_neurons: 1024
@@ -66,45 +66,30 @@ test_dataloader_opts:
batch_size: !ref <batch_size>
num_workers: !ref <batch_size>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -113,6 +98,7 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
diff --git a/recipes/TIMIT/ASR/transducer/hparams/train.yaml b/recipes/TIMIT/ASR/transducer/hparams/train.yaml
index 5b8e53809..204297dc6 100644
--- a/recipes/TIMIT/ASR/transducer/hparams/train.yaml
+++ b/recipes/TIMIT/ASR/transducer/hparams/train.yaml
@@ -28,7 +28,7 @@ data_folder_noise: !ref <data_folder>/noise # The noisy sequencies for data augm
NOISE_DATASET_URL: https://www.dropbox.com/scl/fi/a09pj97s5ifan81dqhi4n/noises.zip?rlkey=j8b0n9kdjdr32o1f06t0cw5b7&dl=1
noise_annotation: !ref <save_folder>/noise.csv #The data manifest files are created by the data preparation script
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 50
batch_size: 8
lr: 1.0
@@ -40,7 +40,7 @@ n_fft: 400
n_mels: 40
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 2
@@ -92,6 +92,8 @@ compute_features: !new:speechbrain.lobes.features.Fbank
normalize: !new:speechbrain.processing.features.InputNormalization
norm_type: global
+############################## Augmentations ###################################
+
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
URL: !ref <NOISE_DATASET_URL>
@@ -99,58 +101,38 @@ prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_U
ext: wav
csv_file: !ref <noise_annotation>
-
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -160,6 +142,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
enc: !new:speechbrain.lobes.models.CRDNN.CRDNN
input_shape: [null, null, !ref <n_mels>]
diff --git a/recipes/TIMIT/ASR/transducer/hparams/train_wav2vec.yaml b/recipes/TIMIT/ASR/transducer/hparams/train_wav2vec.yaml
index 90f899d26..9ead09f56 100644
--- a/recipes/TIMIT/ASR/transducer/hparams/train_wav2vec.yaml
+++ b/recipes/TIMIT/ASR/transducer/hparams/train_wav2vec.yaml
@@ -28,7 +28,7 @@ test_annotation: !ref <save_folder>/test.json
skip_prep: False # Skip data preparation
uppercase: False # Must be True when the TIMIT dataset is in the upper-case version
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 20
batch_size: 8
lr: 0.0003
@@ -41,7 +41,7 @@ sample_rate: 16000
# n_fft: 400
# n_mels: 40
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
# dropout: 0.15
dnn_blocks: 1
@@ -74,45 +74,30 @@ test_dataloader_opts:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
- concat_original: False
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -121,6 +106,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
wav2vec2: !new:speechbrain.lobes.models.huggingface_transformers.wav2vec2.Wav2Vec2
source: !ref <wav2vec2_hub>
output_norm: True
diff --git a/recipes/TIMIT/Alignment/hparams/train.yaml b/recipes/TIMIT/Alignment/hparams/train.yaml
index 7a2a581d9..aaf06b7ff 100644
--- a/recipes/TIMIT/Alignment/hparams/train.yaml
+++ b/recipes/TIMIT/Alignment/hparams/train.yaml
@@ -20,7 +20,7 @@ valid_annotation: !ref <data_folder>/dev.json
test_annotation: !ref <data_folder>/test.json
skip_prep: False # Skip data prep
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 10
batch_size: 256
lr: 0.0003
@@ -40,7 +40,7 @@ phn_set: 60 # {60, 48, 39}
output_neurons: 183
blank_index: 182
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dnn_blocks: 1
dnn_neurons: 2000
diff --git a/recipes/Tedlium2/ASR/transformer/hparams/branchformer_large.yaml b/recipes/Tedlium2/ASR/transformer/hparams/branchformer_large.yaml
index 6ffad0b00..2f9f924c7 100644
--- a/recipes/Tedlium2/ASR/transformer/hparams/branchformer_large.yaml
+++ b/recipes/Tedlium2/ASR/transformer/hparams/branchformer_large.yaml
@@ -34,7 +34,7 @@ valid_csv: !ref <output_folder>/dev/dev.csv
test_csv:
- !ref <output_folder>/test/test.csv
-# Training parameters
+####################### Training Parameters ####################################
# To make Transformers converge, the global bath size should be large enough.
# The global batch size is computed as batch_size * n_gpus * grad_accumulation_factor.
# Empirically, we found that this value should be >= 128.
@@ -97,7 +97,7 @@ valid_dataloader_opts:
test_dataloader_opts:
batch_size: 1
-####################### Model parameters ###########################
+####################### Model Parameters ###########################
# Transformer
d_model: 512
nhead: 8
diff --git a/recipes/Tedlium2/Tokenizer/hparams/tedlium2_500_bpe.yaml b/recipes/Tedlium2/Tokenizer/hparams/tedlium2_500_bpe.yaml
index a97290f13..03c91b126 100644
--- a/recipes/Tedlium2/Tokenizer/hparams/tedlium2_500_bpe.yaml
+++ b/recipes/Tedlium2/Tokenizer/hparams/tedlium2_500_bpe.yaml
@@ -14,7 +14,7 @@ skip_prep: False
train_csv: !ref <output_folder>/train/train.csv
valid_csv: !ref <output_folder>/dev/dev.csv
-# Training parameters
+####################### Training Parameters ####################################
token_type: bpe # ["unigram", "bpe", "char"]
token_output: 500 # index(blank/eos/bos/unk) = 0
character_coverage: 1.0
diff --git a/recipes/UrbanSound8k/SoundClassification/hparams/train_ecapa_tdnn.yaml b/recipes/UrbanSound8k/SoundClassification/hparams/train_ecapa_tdnn.yaml
index 1d9613a16..3ecb1119b 100644
--- a/recipes/UrbanSound8k/SoundClassification/hparams/train_ecapa_tdnn.yaml
+++ b/recipes/UrbanSound8k/SoundClassification/hparams/train_ecapa_tdnn.yaml
@@ -48,7 +48,7 @@ skip_manifest_creation: False
ckpt_interval_minutes: 15 # save checkpoint every N min
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 25
batch_size: 32
lr: 0.001
diff --git a/recipes/Voicebank/ASR/CTC/hparams/train.yaml b/recipes/Voicebank/ASR/CTC/hparams/train.yaml
index 65b833a0b..a49bae5fa 100644
--- a/recipes/Voicebank/ASR/CTC/hparams/train.yaml
+++ b/recipes/Voicebank/ASR/CTC/hparams/train.yaml
@@ -20,7 +20,7 @@ valid_annotation: !ref <output_folder>/valid.json
test_annotation: !ref <output_folder>/test.json
skip_prep: False # Skip data preparation
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 50
batch_size: 8
sorting: ascending
@@ -37,7 +37,7 @@ sample_rate: 16000
n_fft: 400
n_mels: 40
-# Model parameters
+####################### Model Parameters #######################################
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 2
@@ -61,45 +61,31 @@ compute_features: !new:speechbrain.lobes.features.Fbank
n_fft: !ref <n_fft>
n_mels: !ref <n_mels>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 3
max_augmentations: 3
augment_prob: 1.0
@@ -108,6 +94,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
model: !new:speechbrain.lobes.models.CRDNN.CRDNN
input_shape: [null, null, !ref <n_mels>]
activation: !ref <activation>
diff --git a/recipes/Voicebank/MTL/ASR_enhance/hparams/enhance_mimic.yaml b/recipes/Voicebank/MTL/ASR_enhance/hparams/enhance_mimic.yaml
index 2a96e25db..c3391498a 100644
--- a/recipes/Voicebank/MTL/ASR_enhance/hparams/enhance_mimic.yaml
+++ b/recipes/Voicebank/MTL/ASR_enhance/hparams/enhance_mimic.yaml
@@ -18,7 +18,7 @@ valid_annotation: !ref <data_folder>/valid.json
test_annotation: !ref <data_folder>/test.json
skip_prep: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 50
batch_size: 8
lr: 0.0001
diff --git a/recipes/Voicebank/MTL/ASR_enhance/hparams/pretrain_perceptual.yaml b/recipes/Voicebank/MTL/ASR_enhance/hparams/pretrain_perceptual.yaml
index a86fbc4cc..d384d026a 100644
--- a/recipes/Voicebank/MTL/ASR_enhance/hparams/pretrain_perceptual.yaml
+++ b/recipes/Voicebank/MTL/ASR_enhance/hparams/pretrain_perceptual.yaml
@@ -18,7 +18,7 @@ valid_annotation: !ref <data_folder>/valid.json
test_annotation: !ref <data_folder>/test.json
skip_prep: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 20
ctc_epochs: 4
batch_size: 8
diff --git a/recipes/Voicebank/MTL/ASR_enhance/hparams/robust_asr.yaml b/recipes/Voicebank/MTL/ASR_enhance/hparams/robust_asr.yaml
index 1bf087d85..1835342c3 100644
--- a/recipes/Voicebank/MTL/ASR_enhance/hparams/robust_asr.yaml
+++ b/recipes/Voicebank/MTL/ASR_enhance/hparams/robust_asr.yaml
@@ -24,7 +24,7 @@ test_annotation: !ref <data_folder>/test.json
noise_annotation: !ref <save_folder>/noise.csv #The data manifest files are created by the data preparation script
skip_prep: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 30
ctc_epochs: 0
batch_size: 8
@@ -141,6 +141,8 @@ compute_stft: !new:speechbrain.processing.features.STFT
spectral_magnitude: !name:speechbrain.processing.features.spectral_magnitude
power: 0.5
+############################## Augmentations ###################################
+
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
URL: !ref <NOISE_DATASET_URL>
@@ -148,58 +150,38 @@ prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_U
ext: wav
csv_file: !ref <noise_annotation>
-
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -209,7 +191,6 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
-
fbank: !new:speechbrain.lobes.features.Fbank
n_mels: !ref <n_mels>
sample_rate: !ref <sample_rate>
diff --git a/recipes/VoxCeleb/SpeakerRec/hparams/train_ecapa_tdnn.yaml b/recipes/VoxCeleb/SpeakerRec/hparams/train_ecapa_tdnn.yaml
index d7d2be081..c93ba21ec 100644
--- a/recipes/VoxCeleb/SpeakerRec/hparams/train_ecapa_tdnn.yaml
+++ b/recipes/VoxCeleb/SpeakerRec/hparams/train_ecapa_tdnn.yaml
@@ -85,6 +85,7 @@ classifier: !new:speechbrain.lobes.models.ECAPA_TDNN.Classifier
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
+############################## Augmentations ###################################
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
@@ -95,18 +96,14 @@ prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_U
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
-
# Download and prepare the dataset of room impulse responses for augmentation
prepare_rir_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
URL: !ref <RIR_DATASET_URL>
@@ -122,37 +119,24 @@ add_reverb: !new:speechbrain.augment.time_domain.AddReverb
num_workers: !ref <num_workers>
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
parallel_augment: True
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
diff --git a/recipes/VoxCeleb/SpeakerRec/hparams/train_ecapa_tdnn_mel_spec.yaml b/recipes/VoxCeleb/SpeakerRec/hparams/train_ecapa_tdnn_mel_spec.yaml
index 6cb17c6bc..becd8e4d4 100644
--- a/recipes/VoxCeleb/SpeakerRec/hparams/train_ecapa_tdnn_mel_spec.yaml
+++ b/recipes/VoxCeleb/SpeakerRec/hparams/train_ecapa_tdnn_mel_spec.yaml
@@ -104,6 +104,7 @@ classifier: !new:speechbrain.lobes.models.ECAPA_TDNN.Classifier
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
+############################## Augmentations ###################################
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
@@ -114,18 +115,14 @@ prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_U
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
-
# Download and prepare the dataset of room impulse responses for augmentation
prepare_rir_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
URL: !ref <RIR_DATASET_URL>
@@ -141,37 +138,24 @@ add_reverb: !new:speechbrain.augment.time_domain.AddReverb
num_workers: !ref <num_workers>
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
parallel_augment: True
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
diff --git a/recipes/VoxCeleb/SpeakerRec/hparams/train_resnet.yaml b/recipes/VoxCeleb/SpeakerRec/hparams/train_resnet.yaml
index 617457f1a..a20786574 100644
--- a/recipes/VoxCeleb/SpeakerRec/hparams/train_resnet.yaml
+++ b/recipes/VoxCeleb/SpeakerRec/hparams/train_resnet.yaml
@@ -55,7 +55,7 @@ right_frames: 0
deltas: False
# Number of speakers
-# 1211 for vox1, 5994 for vox2, 7205 for vox1+vox2
+# 1211 for vox1, 5994 for vox2, 7205 for vox1+vox2
out_n_neurons: 7205
num_workers: 4
@@ -85,6 +85,7 @@ classifier: !new:speechbrain.lobes.models.ECAPA_TDNN.Classifier
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
+############################## Augmentations ###################################
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
@@ -95,18 +96,14 @@ prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_U
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
-
# Download and prepare the dataset of room impulse responses for augmentation
prepare_rir_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
URL: !ref <RIR_DATASET_URL>
@@ -122,37 +119,24 @@ add_reverb: !new:speechbrain.augment.time_domain.AddReverb
num_workers: !ref <num_workers>
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
parallel_augment: True
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
diff --git a/recipes/VoxCeleb/SpeakerRec/hparams/train_x_vectors.yaml b/recipes/VoxCeleb/SpeakerRec/hparams/train_x_vectors.yaml
index 8a70462c5..ab628c681 100644
--- a/recipes/VoxCeleb/SpeakerRec/hparams/train_x_vectors.yaml
+++ b/recipes/VoxCeleb/SpeakerRec/hparams/train_x_vectors.yaml
@@ -88,6 +88,7 @@ classifier: !new:speechbrain.lobes.models.Xvector.Classifier
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
+############################## Augmentations ###################################
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
@@ -98,18 +99,14 @@ prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_U
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
-
# Download and prepare the dataset of room impulse responses for augmentation
prepare_rir_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
URL: !ref <RIR_DATASET_URL>
@@ -125,37 +122,24 @@ add_reverb: !new:speechbrain.augment.time_domain.AddReverb
num_workers: !ref <num_workers>
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
parallel_augment: True
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
diff --git a/recipes/VoxLingua107/lang_id/hparams/train_ecapa.yaml b/recipes/VoxLingua107/lang_id/hparams/train_ecapa.yaml
index a2d872486..db29a301b 100644
--- a/recipes/VoxLingua107/lang_id/hparams/train_ecapa.yaml
+++ b/recipes/VoxLingua107/lang_id/hparams/train_ecapa.yaml
@@ -58,6 +58,8 @@ val_dataloader_options:
num_workers: 1
batch_size: !ref <batch_size_val>
+############################## Augmentations ###################################
+
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
URL: !ref <NOISE_DATASET_URL>
@@ -72,7 +74,6 @@ prepare_rir_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
ext: wav
csv_file: !ref <rir_annotation>
-
# Add reverberation to input signal
add_reverb: !new:speechbrain.augment.time_domain.AddReverb
csv_file: !ref <rir_annotation>
@@ -81,27 +82,21 @@ add_reverb: !new:speechbrain.augment.time_domain.AddReverb
num_workers: !ref <num_workers>
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
# Speed perturbation
-speed_changes: [90, 100, 110] # List of speed changes for time-stretching
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
shuffle_augmentations: True
min_augmentations: 1
max_augmentations: 3
diff --git a/recipes/WHAMandWHAMR/enhancement/hparams/cnntransformer-wham-DM.yaml b/recipes/WHAMandWHAMR/enhancement/hparams/cnntransformer-wham-DM.yaml
index a38bb8c88..843c9fb09 100644
--- a/recipes/WHAMandWHAMR/enhancement/hparams/cnntransformer-wham-DM.yaml
+++ b/recipes/WHAMandWHAMR/enhancement/hparams/cnntransformer-wham-DM.yaml
@@ -45,7 +45,7 @@ save_audio: True # Save estimated sources on disk
sample_rate: 8000
n_audio_to_save: 20
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 8
lr: 0.0001
diff --git a/recipes/WHAMandWHAMR/enhancement/hparams/cnntransformer-whamr-DM.yaml b/recipes/WHAMandWHAMR/enhancement/hparams/cnntransformer-whamr-DM.yaml
index af3fcb0d7..b55f05a11 100644
--- a/recipes/WHAMandWHAMR/enhancement/hparams/cnntransformer-whamr-DM.yaml
+++ b/recipes/WHAMandWHAMR/enhancement/hparams/cnntransformer-whamr-DM.yaml
@@ -45,7 +45,7 @@ save_audio: True # Save estimated sources on disk
sample_rate: 8000
n_audio_to_save: 20
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 8
lr: 0.0001
diff --git a/recipes/WHAMandWHAMR/enhancement/hparams/convtasnet-whamr-DM.yaml b/recipes/WHAMandWHAMR/enhancement/hparams/convtasnet-whamr-DM.yaml
index 7c2990442..545428d5a 100644
--- a/recipes/WHAMandWHAMR/enhancement/hparams/convtasnet-whamr-DM.yaml
+++ b/recipes/WHAMandWHAMR/enhancement/hparams/convtasnet-whamr-DM.yaml
@@ -44,7 +44,7 @@ save_audio: True # Save estimated sources on disk
sample_rate: 8000
n_audio_to_save: 20
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 10
lr: 0.00015
diff --git a/recipes/WHAMandWHAMR/enhancement/hparams/dprnn-whamr-DM.yaml b/recipes/WHAMandWHAMR/enhancement/hparams/dprnn-whamr-DM.yaml
index f3158a625..d974b03c1 100644
--- a/recipes/WHAMandWHAMR/enhancement/hparams/dprnn-whamr-DM.yaml
+++ b/recipes/WHAMandWHAMR/enhancement/hparams/dprnn-whamr-DM.yaml
@@ -44,7 +44,7 @@ save_audio: True # Save estimated sources on disk
sample_rate: 8000
n_audio_to_save: 20
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-wham.yaml b/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-wham.yaml
index 75c90a0f1..df1935306 100644
--- a/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-wham.yaml
+++ b/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-wham.yaml
@@ -44,7 +44,7 @@ noprogressbar: False
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr-16k-DM.yaml b/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr-16k-DM.yaml
index 625801e51..dc2783491 100644
--- a/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr-16k-DM.yaml
+++ b/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr-16k-DM.yaml
@@ -46,7 +46,7 @@ save_audio: True # Save estimated sources on disk
sample_rate: 16000
n_audio_to_save: 20
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr-16k.yaml b/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr-16k.yaml
index d42060642..d11332c7e 100644
--- a/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr-16k.yaml
+++ b/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr-16k.yaml
@@ -44,7 +44,7 @@ save_audio: True # Save estimated sources on disk
sample_rate: 16000
n_audio_to_save: 20
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr-DM.yaml b/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr-DM.yaml
index 2acd34370..3721698bc 100644
--- a/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr-DM.yaml
+++ b/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr-DM.yaml
@@ -44,7 +44,7 @@ save_audio: True # Save estimated sources on disk
sample_rate: 8000
n_audio_to_save: 20
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr.yaml b/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr.yaml
index 536a46f49..14a38a06b 100644
--- a/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr.yaml
+++ b/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-whamr.yaml
@@ -44,7 +44,7 @@ save_audio: True # Save estimated sources on disk
sample_rate: 8000
n_audio_to_save: 20
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WHAMandWHAMR/separation/hparams/sepformer-wham.yaml b/recipes/WHAMandWHAMR/separation/hparams/sepformer-wham.yaml
index 1661d68f2..db920e7fa 100644
--- a/recipes/WHAMandWHAMR/separation/hparams/sepformer-wham.yaml
+++ b/recipes/WHAMandWHAMR/separation/hparams/sepformer-wham.yaml
@@ -42,7 +42,7 @@ noprogressbar: False
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WHAMandWHAMR/separation/hparams/sepformer-whamr.yaml b/recipes/WHAMandWHAMR/separation/hparams/sepformer-whamr.yaml
index d4bc250d0..8538529a6 100644
--- a/recipes/WHAMandWHAMR/separation/hparams/sepformer-whamr.yaml
+++ b/recipes/WHAMandWHAMR/separation/hparams/sepformer-whamr.yaml
@@ -40,7 +40,7 @@ num_spks: 2 # set to 3 for wsj0-3mix
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WSJ0Mix/separation/hparams/convtasnet.yaml b/recipes/WSJ0Mix/separation/hparams/convtasnet.yaml
index 038330504..0305c6236 100644
--- a/recipes/WSJ0Mix/separation/hparams/convtasnet.yaml
+++ b/recipes/WSJ0Mix/separation/hparams/convtasnet.yaml
@@ -36,7 +36,7 @@ noprogressbar: False
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WSJ0Mix/separation/hparams/dprnn.yaml b/recipes/WSJ0Mix/separation/hparams/dprnn.yaml
index a78a78266..df1952d8c 100644
--- a/recipes/WSJ0Mix/separation/hparams/dprnn.yaml
+++ b/recipes/WSJ0Mix/separation/hparams/dprnn.yaml
@@ -36,7 +36,7 @@ noprogressbar: False
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WSJ0Mix/separation/hparams/resepformer.yaml b/recipes/WSJ0Mix/separation/hparams/resepformer.yaml
index 2b2711f52..406f2aa76 100644
--- a/recipes/WSJ0Mix/separation/hparams/resepformer.yaml
+++ b/recipes/WSJ0Mix/separation/hparams/resepformer.yaml
@@ -37,7 +37,7 @@ num_spks: 2 # set to 3 for wsj0-3mix
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WSJ0Mix/separation/hparams/sepformer-conformerintra.yaml b/recipes/WSJ0Mix/separation/hparams/sepformer-conformerintra.yaml
index 15550c9a2..2cf2b7ac5 100644
--- a/recipes/WSJ0Mix/separation/hparams/sepformer-conformerintra.yaml
+++ b/recipes/WSJ0Mix/separation/hparams/sepformer-conformerintra.yaml
@@ -37,7 +37,7 @@ num_spks: 2 # set to 3 for wsj0-3mix
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WSJ0Mix/separation/hparams/sepformer-customdataset.yaml b/recipes/WSJ0Mix/separation/hparams/sepformer-customdataset.yaml
index 82a2d3009..c896f2dfd 100644
--- a/recipes/WSJ0Mix/separation/hparams/sepformer-customdataset.yaml
+++ b/recipes/WSJ0Mix/separation/hparams/sepformer-customdataset.yaml
@@ -39,7 +39,7 @@ noprogressbar: False
save_audio: True # Save estimated sources on disk
sample_rate: 16000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WSJ0Mix/separation/hparams/sepformer.yaml b/recipes/WSJ0Mix/separation/hparams/sepformer.yaml
index 4787fb3aa..77319604d 100644
--- a/recipes/WSJ0Mix/separation/hparams/sepformer.yaml
+++ b/recipes/WSJ0Mix/separation/hparams/sepformer.yaml
@@ -40,7 +40,7 @@ save_audio: True # Save estimated sources on disk
n_audio_to_save: 20
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/WSJ0Mix/separation/hparams/skim.yaml b/recipes/WSJ0Mix/separation/hparams/skim.yaml
index 53b312efd..606c7060a 100644
--- a/recipes/WSJ0Mix/separation/hparams/skim.yaml
+++ b/recipes/WSJ0Mix/separation/hparams/skim.yaml
@@ -37,7 +37,7 @@ num_spks: 2 # set to 3 for wsj0-3mix
save_audio: False # Save estimated sources on disk
sample_rate: 8000
-# Training parameters
+####################### Training Parameters ####################################
N_epochs: 200
batch_size: 1
lr: 0.00015
diff --git a/recipes/ZaionEmotionDataset/emotion_diarization/hparams/train.yaml b/recipes/ZaionEmotionDataset/emotion_diarization/hparams/train.yaml
index 644ca89a2..0d9601d20 100644
--- a/recipes/ZaionEmotionDataset/emotion_diarization/hparams/train.yaml
+++ b/recipes/ZaionEmotionDataset/emotion_diarization/hparams/train.yaml
@@ -30,7 +30,7 @@ train_annotation: !ref <output_folder>/train.json
valid_annotation: !ref <output_folder>/valid.json
test_annotation: !ref <output_folder>/test.json
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 15
lr: 0.0001
lr_wav2vec: 0.00001
diff --git a/recipes/fluent-speech-commands/Tokenizer/hparams/tokenizer_bpe51.yaml b/recipes/fluent-speech-commands/Tokenizer/hparams/tokenizer_bpe51.yaml
index db7c1ddb7..eff38c7bf 100644
--- a/recipes/fluent-speech-commands/Tokenizer/hparams/tokenizer_bpe51.yaml
+++ b/recipes/fluent-speech-commands/Tokenizer/hparams/tokenizer_bpe51.yaml
@@ -13,7 +13,7 @@ train_csv: !ref <output_folder>/train.csv
valid_csv: !ref <output_folder>/valid.csv
skip_prep: False
-# Training parameters
+####################### Training Parameters ####################################
token_type: unigram # ["unigram", "bpe", "char"]
token_output: 51 # index(blank/eos/bos/unk) = 0
character_coverage: 1.0
diff --git a/recipes/fluent-speech-commands/direct/hparams/train.yaml b/recipes/fluent-speech-commands/direct/hparams/train.yaml
index a7c072343..428faf144 100644
--- a/recipes/fluent-speech-commands/direct/hparams/train.yaml
+++ b/recipes/fluent-speech-commands/direct/hparams/train.yaml
@@ -32,14 +32,14 @@ rir_annotation: !ref <save_folder>/rir.csv
tokenizer_file: https://www.dropbox.com/s/hvf2huofnq0sjbn/51_unigram.model?dl=1
skip_prep: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 6
batch_size: 16
lr: 0.0003
# token_type: unigram # ["unigram", "bpe", "char"]
sorting: random
-# Model parameters
+####################### Model Parameters #######################################
sample_rate: 16000
emb_size: 128
dec_neurons: 512
@@ -65,6 +65,8 @@ dataloader_opts:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
+############################## Augmentations ###################################
+
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
URL: !ref <NOISE_DATASET_URL>
@@ -87,45 +89,32 @@ add_reverb: !new:speechbrain.augment.time_domain.AddReverb
num_workers: !ref <num_workers>
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 9
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 3 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 2
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
shuffle_augmentations: True
min_augmentations: 1
max_augmentations: 4
@@ -136,7 +125,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
-# Models
+############################## Models ##########################################
+
asr_model_source: speechbrain/asr-crdnn-rnnlm-librispeech
slu_enc: !new:speechbrain.nnet.containers.Sequential
diff --git a/recipes/timers-and-such/LM/hparams/train.yaml b/recipes/timers-and-such/LM/hparams/train.yaml
index 485dd5426..f3ba652ed 100644
--- a/recipes/timers-and-such/LM/hparams/train.yaml
+++ b/recipes/timers-and-such/LM/hparams/train.yaml
@@ -23,7 +23,7 @@ csv_test_synth: !ref <output_folder>/test-synth-type=decoupled.csv
csv_test_real: !ref <output_folder>/test-real-type=decoupled.csv
skip_prep: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 10
batch_size: 128
lr: 0.0003
diff --git a/recipes/timers-and-such/Tokenizer/hparams/tokenizer_bpe51.yaml b/recipes/timers-and-such/Tokenizer/hparams/tokenizer_bpe51.yaml
index 7554a0342..2a9f39161 100644
--- a/recipes/timers-and-such/Tokenizer/hparams/tokenizer_bpe51.yaml
+++ b/recipes/timers-and-such/Tokenizer/hparams/tokenizer_bpe51.yaml
@@ -15,7 +15,7 @@ train_csv: !ref <output_folder>/train-type=direct.csv
valid_csv: !ref <output_folder>/dev-real-type=direct.csv
-# Training parameters
+####################### Training Parameters ####################################
token_type: unigram # ["unigram", "bpe", "char"]
token_output: 51 # index(blank/eos/bos/unk) = 0
character_coverage: 1.0
diff --git a/recipes/timers-and-such/decoupled/hparams/train_LS_LM.yaml b/recipes/timers-and-such/decoupled/hparams/train_LS_LM.yaml
index 212752668..1ee56d561 100644
--- a/recipes/timers-and-such/decoupled/hparams/train_LS_LM.yaml
+++ b/recipes/timers-and-such/decoupled/hparams/train_LS_LM.yaml
@@ -34,7 +34,7 @@ skip_prep: False
ckpt_interval_minutes: 15 # save checkpoint every N min
test_on_all_real: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 1
batch_size: 16
lr: 0.0003
diff --git a/recipes/timers-and-such/decoupled/hparams/train_TAS_LM.yaml b/recipes/timers-and-such/decoupled/hparams/train_TAS_LM.yaml
index fcf6393e6..5f0d93d09 100644
--- a/recipes/timers-and-such/decoupled/hparams/train_TAS_LM.yaml
+++ b/recipes/timers-and-such/decoupled/hparams/train_TAS_LM.yaml
@@ -34,7 +34,7 @@ skip_prep: False
ckpt_interval_minutes: 15 # save checkpoint every N min
test_on_all_real: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 1
batch_size: 16
lr: 0.0003
diff --git a/recipes/timers-and-such/direct/hparams/train.yaml b/recipes/timers-and-such/direct/hparams/train.yaml
index 4fb574fc3..01909eb5b 100644
--- a/recipes/timers-and-such/direct/hparams/train.yaml
+++ b/recipes/timers-and-such/direct/hparams/train.yaml
@@ -38,14 +38,14 @@ data_folder_noise: !ref <data_folder>/noise # The noisy sequencies for data augm
NOISE_DATASET_URL: https://www.dropbox.com/scl/fi/a09pj97s5ifan81dqhi4n/noises.zip?rlkey=j8b0n9kdjdr32o1f06t0cw5b7&dl=1
noise_annotation: !ref <save_folder>/noise.csv #The data manifest files are created by the data preparation script
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 1
batch_size: 16
lr: 0.0003
# token_type: unigram # ["unigram", "bpe", "char"]
sorting: random
-# Model parameters
+####################### Model Parameters #######################################
sample_rate: 16000
emb_size: 128
dec_neurons: 512
@@ -71,6 +71,7 @@ dataloader_opts:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
+############################## Augmentations ###################################
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
@@ -80,56 +81,37 @@ prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_U
csv_file: !ref <noise_annotation>
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -139,8 +121,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
-# Models
asr_model_source: speechbrain/asr-crdnn-rnnlm-librispeech
slu_enc: !new:speechbrain.nnet.containers.Sequential
diff --git a/recipes/timers-and-such/direct/hparams/train_with_wav2vec2.yaml b/recipes/timers-and-such/direct/hparams/train_with_wav2vec2.yaml
index b9b451e91..b9ad3cfc2 100644
--- a/recipes/timers-and-such/direct/hparams/train_with_wav2vec2.yaml
+++ b/recipes/timers-and-such/direct/hparams/train_with_wav2vec2.yaml
@@ -37,7 +37,7 @@ ckpt_interval_minutes: 15 # save checkpoint every N min
test_on_all_real: False
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 50
batch_size: 8
lr: 0.0004
@@ -49,7 +49,7 @@ freeze_wav2vec: False
# token_type: unigram # ["unigram", "bpe", "char"]
sorting: ascending
-# Model parameters
+####################### Model Parameters #######################################
sample_rate: 16000
emb_size: 128
dec_neurons: 512
@@ -171,45 +171,31 @@ checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
lr_annealing_wav2vec2: !ref <lr_annealing_wav2vec2>
counter: !ref <epoch_counter>
-# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
+############################## Augmentations ###################################
+# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
diff --git a/recipes/timers-and-such/multistage/hparams/train_LS_LM.yaml b/recipes/timers-and-such/multistage/hparams/train_LS_LM.yaml
index 3d1c4e156..b804df9f9 100644
--- a/recipes/timers-and-such/multistage/hparams/train_LS_LM.yaml
+++ b/recipes/timers-and-such/multistage/hparams/train_LS_LM.yaml
@@ -39,7 +39,7 @@ NOISE_DATASET_URL: https://www.dropbox.com/scl/fi/a09pj97s5ifan81dqhi4n/noises.z
noise_annotation: !ref <save_folder>/noise.csv #The data manifest files are created by the data preparation script
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 1
batch_size: 16
lr: 0.0003
@@ -73,6 +73,8 @@ dataloader_opts:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
+############################## Augmentations ###################################
+
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
URL: !ref <NOISE_DATASET_URL>
@@ -81,57 +83,37 @@ prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_U
csv_file: !ref <noise_annotation>
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
-# Augmenter: Combines previously defined augmentations to perform data augmentation
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -141,7 +123,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
-# Models
+############################## Models ##########################################
+
asr_model: !apply:speechbrain.inference.ASR.EncoderDecoderASR.from_hparams
source: speechbrain/asr-crdnn-rnnlm-librispeech
run_opts: {"device":"cuda:0"}
diff --git a/recipes/timers-and-such/multistage/hparams/train_TAS_LM.yaml b/recipes/timers-and-such/multistage/hparams/train_TAS_LM.yaml
index d21f309c5..56eb59d20 100644
--- a/recipes/timers-and-such/multistage/hparams/train_TAS_LM.yaml
+++ b/recipes/timers-and-such/multistage/hparams/train_TAS_LM.yaml
@@ -39,7 +39,7 @@ NOISE_DATASET_URL: https://www.dropbox.com/scl/fi/a09pj97s5ifan81dqhi4n/noises.z
noise_annotation: !ref <save_folder>/noise.csv #The data manifest files are created by the data preparation script
-# Training parameters
+####################### Training Parameters ####################################
number_of_epochs: 1
batch_size: 16
lr: 0.0003
@@ -73,6 +73,8 @@ dataloader_opts:
epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
limit: !ref <number_of_epochs>
+############################## Augmentations ###################################
+
# Download and prepare the dataset of noisy sequences for augmentation
prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_URL
URL: !ref <NOISE_DATASET_URL>
@@ -81,57 +83,37 @@ prepare_noise_data: !name:speechbrain.augment.preparation.prepare_dataset_from_U
csv_file: !ref <noise_annotation>
# Add noise to input signal
-snr_low: 0 # Min SNR for noise augmentation
-snr_high: 15 # Max SNR for noise augmentation
-
add_noise: !new:speechbrain.augment.time_domain.AddNoise
csv_file: !ref <noise_annotation>
- snr_low: !ref <snr_low>
- snr_high: !ref <snr_high>
+ snr_low: 0
+ snr_high: 15
noise_sample_rate: !ref <sample_rate>
clean_sample_rate: !ref <sample_rate>
num_workers: !ref <num_workers>
# Speed perturbation
-speed_changes: [95, 100, 105] # List of speed changes for time-stretching
-
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
orig_freq: !ref <sample_rate>
- speeds: !ref <speed_changes>
+ speeds: [95, 100, 105]
# Frequency drop: randomly drops a number of frequency bands to zero.
-drop_freq_low: 0 # Min frequency band dropout probability
-drop_freq_high: 1 # Max frequency band dropout probability
-drop_freq_count_low: 1 # Min number of frequency bands to drop
-drop_freq_count_high: 3 # Max number of frequency bands to drop
-drop_freq_width: 0.05 # Width of frequency bands to drop
-
drop_freq: !new:speechbrain.augment.time_domain.DropFreq
- drop_freq_low: !ref <drop_freq_low>
- drop_freq_high: !ref <drop_freq_high>
- drop_freq_count_low: !ref <drop_freq_count_low>
- drop_freq_count_high: !ref <drop_freq_count_high>
- drop_freq_width: !ref <drop_freq_width>
+ drop_freq_low: 0
+ drop_freq_high: 1
+ drop_freq_count_low: 1
+ drop_freq_count_high: 3
+ drop_freq_width: 0.05
# Time drop: randomly drops a number of temporal chunks.
-drop_chunk_count_low: 1 # Min number of audio chunks to drop
-drop_chunk_count_high: 5 # Max number of audio chunks to drop
-drop_chunk_length_low: 1000 # Min length of audio chunks to drop
-drop_chunk_length_high: 2000 # Max length of audio chunks to drop
-
drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
- drop_length_low: !ref <drop_chunk_length_low>
- drop_length_high: !ref <drop_chunk_length_high>
- drop_count_low: !ref <drop_chunk_count_low>
- drop_count_high: !ref <drop_chunk_count_high>
+ drop_length_low: 1000
+ drop_length_high: 2000
+ drop_count_low: 1
+ drop_count_high: 5
-# Augmenter: Combines previously defined augmentations to perform data augmentation
# Augmenter: Combines previously defined augmentations to perform data augmentation
wav_augment: !new:speechbrain.augment.augmenter.Augmenter
- parallel_augment: False
concat_original: True
- repeat_augment: 1
- shuffle_augmentations: False
min_augmentations: 4
max_augmentations: 4
augment_prob: 1.0
@@ -141,6 +123,8 @@ wav_augment: !new:speechbrain.augment.augmenter.Augmenter
!ref <drop_freq>,
!ref <drop_chunk>]
+############################## Models ##########################################
+
# Models
asr_model: !apply:speechbrain.inference.ASR.EncoderDecoderASR.from_hparams
source: speechbrain/asr-crdnn-rnnlm-librispeech
diff --git a/speechbrain/augment/augmenter.py b/speechbrain/augment/augmenter.py
index 25ea17480..55aee785c 100644
--- a/speechbrain/augment/augmenter.py
+++ b/speechbrain/augment/augmenter.py
@@ -438,20 +438,52 @@ class Augmenter(torch.nn.Module):
return output, output_lengths
+ def replicate_multiple_labels(self, *args):
+ """
+ Replicates the labels along the batch axis a number of times that
+ corresponds to the number of augmentations. Indeed parallel and
+ concatenation augmentations alter the time dimension.
+
+ Arguments
+ ---------
+ args : torch.Tensor
+ Input label tensors to be replicated. Can be a uniq or a list of
+ Tensors.
+
+ Returns
+ -------
+ augmented_labels: torch.Tensor
+ Labels corresponding to the augmented input. Returns as many Tensor
+ as given in input.
+ """
+
+ # Determine whether to apply data augmentation
+ if not self.do_augment:
+ return args
+
+ list_of_augmented_labels = []
+
+ for labels in args:
+ list_of_augmented_labels.append(self.replicate_labels(labels))
+
+ return list_of_augmented_labels
+
def replicate_labels(self, labels):
"""
Replicates the labels along the batch axis a number of times that
- corresponds to the number of augmentations.
+ corresponds to the number of augmentations. Indeed parallel and
+ concatenation augmentations alter the time dimension.
Arguments
---------
labels : torch.Tensor
- Input label tensor to be replicated.
+ Input label tensors to be replicated.
Returns
-------
augmented_labels: torch.Tensor
- Labels corresponding to the augmented input.
+ Labels corresponding to the augmented input. Returns as many Tensor
+ as given in input.
"""
# Determine whether to apply data augmentation
@@ -477,6 +509,7 @@ class Augmenter(torch.nn.Module):
)
augmented_labels = torch.cat(augmented_labels, dim=0)
+
return augmented_labels
def check_min_max_augmentations(self):
--
GitLab