deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-05 18:09:24 +08:00
Code Issues Packages Projects Releases Wiki Activity

[debug] a better solution for mismatch of speech feat len and speech token len, refer to https://github.com/FunAudioLLM/CosyVoice/issues/1051

Browse Source
This commit is contained in:
burkliu
2025-04-24 17:14:49 +08:00
parent 587604b2b4
commit 65ad448714
2 changed files with 10 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
cosyvoice/dataset/processor.py
View File

@@ -159,6 +159,7 @@ def truncate(data, truncate_length=24576, mode='train'):
def compute_fbank(data,
feat_extractor,
token_mel_ratio=2,
mode='train'):
""" Extract fbank
@@ -174,8 +175,15 @@ def compute_fbank(data,
assert 'utt' in sample
assert 'text_token' in sample
waveform = sample['speech']
mat = feat_extractor(waveform).squeeze(dim=0).transpose(0, 1)
sample['speech_feat'] = mat
feat = feat_extractor(waveform).squeeze(dim=0).transpose(0, 1)
# padding with replicate mode (align to speech_token len * token_mel_ratio)
pad_len = sample["speech_token"].shape[0] * token_mel_ratio - feat.shape[0]
if pad_len > 0:
feat_to_pad = feat[-1:].repeat((pad_len, 1))
feat = torch.cat([feat, feat_to_pad], dim=0)
sample['speech_feat'] = feat
yield sample