deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-04 09:29:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

[feature] modify pad to trim

Browse Source
This commit is contained in:
burkliu
2025-04-25 10:31:43 +08:00
parent 65ad448714
commit 038ff9f353
1 changed files with 4 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
cosyvoice/dataset/processor.py
View File

@@ -177,11 +177,10 @@ def compute_fbank(data,
waveform = sample['speech']
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)
# 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