fixes

recipes/Voicebank/enhance/MetricGAN-U/hparams/train_dnsmos.yaml

@@ -14,7 +14,7 @@ __set_seed: !!python/object/apply:torch.manual_seed [!ref <seed>]
 data_folder: /scratch/jo15g23/vbd/noisy-vctk-16k # e.g, /data/member1/user_jasonfu/noisy-vctk-16k
 
 MetricGAN_folder: !ref <output_folder>/enhanced_wavs
-output_folder: !ref /scratch/jo15g23/metricgan/results/<seed>
+output_folder: !ref /scratch/jo15g23/metricgan_new/<seed>
 save_folder: !ref <output_folder>/save
 train_log: !ref <output_folder>/train_log.txt
 enhanced_folder: !ref <output_folder>/enhanced_wavs

recipes/Voicebank/enhance/MetricGAN-U/train.py

@@ -32,8 +32,7 @@ from speechbrain.nnet.loss.stoi_loss import stoi_loss
 from speechbrain.utils.distributed import run_on_main
 from speechbrain.dataio.sampler import ReproducibleWeightedRandomSampler
 
-
-from seframework.utils.dnsmos import DNSMOS
+from speechmos import dnsmos as dnsmos_func
 
 ### For DNSMSOS
 # URL for the web service
@@ -49,8 +48,6 @@ headers = {"Content-Type": "application/json"}
 # If authentication is enabled, set the authorization header
 headers["Authorization"] = f"Basic {AUTH_KEY }"
 
-dnsmos= DNSMOS(device_id=0, sample_rate=16000, cache_session=True, p835=False)
-
 def sigmoid(x):
     s = 1 / (1 + np.exp(-x))
     return s
@@ -112,8 +109,8 @@ def dnsmos_eval(predict, target):
     pred_wav = pred_wav / max(abs(pred_wav))
     pred_wav = pred_wav[np.newaxis, :]
     
-    scores = dnsmos(pred_wav)
-    score = scores["mos"]
+    scores = dnsmos_func.run(pred_wav, sr=16000)
+    score = scores["p808_mos"]
     score = float(sigmoid(score))  # normalize the score to 0~1
     return score
 
@@ -148,8 +145,8 @@ def dnsmos_eval_valid(predict, target):
     pred_wav = pred_wav.numpy()
     pred_wav = pred_wav / max(abs(pred_wav))
     
-    scores = dnsmos(pred_wav)
-    score = scores["mos"]
+    scores = dnsmos_func.run(pred_wav, sr=16000)
+    score = scores["p808_mos"]
     return score
 
     # data = {"data": pred_wav.tolist()}