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

Merge pull request #1232 from boji123/bj_dev_feat_len_pad

Browse Source 
a better solution for mismatch of speech feat len and speech token len when trainning
This commit is contained in:
Xiang Lyu
2025-04-30 09:41:50 +08:00
committed by GitHub
parent a442317d17 038ff9f353
commit 88f467a8ac
6 changed files with 59 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
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 +175,14 @@ 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)
# trim to align speech_token and speech_feat
token_len = min(feat.shape[0] // token_mel_ratio, sample["speech_token"].shape[0])
feat = feat[:token_mel_ratio * token_len]
sample["speech_token"] = sample["speech_token"][:token_len]
sample['speech_feat'] = feat
yield sample