From 28d0b9f31d54e7c8c99bb3361fbedac80241dc8c Mon Sep 17 00:00:00 2001
From: Gaspard Petit <gaspardpetit@gmail.com>
Date: Sun, 4 Feb 2024 17:45:08 -0500
Subject: [PATCH] Fix for issue 2334 : VAD fails on short audio (#2335)
* Fix for issue 2334 : VAD fails on short audio
* Fix off-by-one boundary logic in VAD processing
---
speechbrain/inference/VAD.py | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/speechbrain/inference/VAD.py b/speechbrain/inference/VAD.py
index 6ba7db006..6f3d55e9c 100644
--- a/speechbrain/inference/VAD.py
+++ b/speechbrain/inference/VAD.py
@@ -109,6 +109,10 @@ class VAD(Pretrained):
last_chunk = False
begin_sample = 0
while True:
+ # Check if the current chunk is the last one
+ if begin_sample + long_chunk_len >= audio_len:
+ last_chunk = True
+
# Reading the big chunk
large_chunk, fs = torchaudio.load(
str(audio_file),
@@ -171,10 +175,6 @@ class VAD(Pretrained):
# Update counter to process the next big chunk
begin_sample = begin_sample + long_chunk_len
- # Check if the current chunk is the last one
- if begin_sample + long_chunk_len > audio_len:
- last_chunk = True
-
# Converting the list to a tensor
prob_vad = torch.cat(prob_chunks, dim=1)
last_elem = int(audio_len / (self.time_resolution * sample_rate))
--
GitLab