optimize resampling

Author: yxlllc

process.py

@@ -38,7 +38,10 @@ def wav2spec(warp):
         audio, sr = torchaudio.load(pathslist[0])
         if sr != config['audio_sample_rate']:
             if sr > config['audio_sample_rate']:
-                audio = torchaudio.transforms.Resample(orig_freq=sr, new_freq=config['audio_sample_rate'])(audio)
+                audio = torchaudio.transforms.Resample(
+                                orig_freq=sr, 
+                                new_freq=config['audio_sample_rate'],
+                                lowpass_filter_width=128)(audio)
             else:
             # audio= torchaudio.transforms.Resample(orig_freq=sr, new_freq=config['audio_sample_rate'])(audio)
                 print('error:flie_', str(pathslist[0]),

process_aug.py

@@ -25,7 +25,10 @@ def dynamic_range_compression_torch(x, C=1, clip_val=1e-9):
 def wav_aug(wav, hop_size, speed=1):
     orig_freq = int(np.round(hop_size * speed))
     new_freq = hop_size
-    return torchaudio.transforms.Resample(orig_freq=orig_freq, new_freq=new_freq)(wav)
+    return torchaudio.transforms.Resample(
+                orig_freq=orig_freq, 
+                new_freq=new_freq, 
+                lowpass_filter_width=128)(wav)
 
 
 def wav2spec(warp):

training/nsf_HiFigan_task.py

@@ -43,7 +43,10 @@ def dynamic_range_compression_torch(x, C=1, clip_val=1e-9):
 def wav_aug(wav, hop_size, speed=1):
     orig_freq = int(np.round(hop_size * speed))
     new_freq = hop_size
-    return torchaudio.transforms.Resample(orig_freq=orig_freq, new_freq=new_freq)(wav)
+    return torchaudio.transforms.Resample(
+                orig_freq=orig_freq, 
+                new_freq=new_freq, 
+                lowpass_filter_width=128)(wav)
 
 
 class nsf_HiFigan_dataset(Dataset):