mrq/DL-Art-School
1
3
Fork 5
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fix nan issue when interpolating audio

Browse Source
This commit is contained in:
James Betker 2021-08-14 20:42:01 -06:00
parent 769f0acc53
commit ad3391bd96
2 changed files with 9 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
codes/data/audio/nv_tacotron_dataset.py
View File

@ -79,20 +79,22 @@ class TextMelLoader(torch.utils.data.Dataset):
if not self.load_mel_from_disk: if not self.load_mel_from_disk:
if filename.endswith('.wav'): if filename.endswith('.wav'):
audio, sampling_rate = load_wav_to_torch(filename) audio, sampling_rate = load_wav_to_torch(filename)
audio = (audio / self.max_wav_value).clip(-1,1) audio = (audio / self.max_wav_value)
else: else:
audio, sampling_rate = audio2numpy.audio_from_file(filename) audio, sampling_rate = audio2numpy.audio_from_file(filename)
audio = torch.tensor(audio) audio = torch.tensor(audio)
audio = (audio.squeeze().clip(-1,1))
if sampling_rate != self.input_sample_rate: if sampling_rate != self.input_sample_rate:
if sampling_rate < self.input_sample_rate: if sampling_rate < self.input_sample_rate:
print(f'{filename} has a sample rate of {sampling_rate} which is lower than the requested sample rate of {self.input_sample_rate}. This is not a good idea.') print(f'{filename} has a sample rate of {sampling_rate} which is lower than the requested sample rate of {self.input_sample_rate}. This is not a good idea.')
audio = torch.nn.functional.interpolate(audio.unsqueeze(0).unsqueeze(1), scale_factor=self.input_sample_rate/sampling_rate, mode='area', recompute_scale_factor=False).squeeze() audio_norm = torch.nn.functional.interpolate(audio.unsqueeze(0).unsqueeze(1), scale_factor=self.input_sample_rate/sampling_rate, mode='nearest', recompute_scale_factor=False).squeeze()
if (audio.min() < -1).any() or (audio.max() > 1).any(): else:
print(f"Error with audio ranging for {filename}; min={audio.min()} max={audio.max()}") audio_norm = audio
if audio_norm.std() > 1:
print(f"Something is very wrong with the given audio. std_dev={audio_norm.std()}. file={filename}")
return None return None
audio_norm = audio.unsqueeze(0) audio_norm.clip_(-1, 1)
audio_norm = audio_norm.unsqueeze(0)
audio_norm = torch.autograd.Variable(audio_norm, requires_grad=False) audio_norm = torch.autograd.Variable(audio_norm, requires_grad=False)
if self.input_sample_rate != self.sampling_rate: if self.input_sample_rate != self.sampling_rate:
ratio = self.sampling_rate / self.input_sample_rate ratio = self.sampling_rate / self.input_sample_rate

2
codes/models/tacotron2/text/__init__.py
View File

@ -76,4 +76,4 @@ def _arpabet_to_sequence(text):
def _should_keep_symbol(s): def _should_keep_symbol(s):
return s in _symbol_to_id and s is not '_' and s is not '~' return s in _symbol_to_id and s != '_' and s != '~'