deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-05 01:49:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #408 from FunAudioLLM/dev/lyuxiang.lx

Browse Source 
Dev/lyuxiang.lx
This commit is contained in:
Xiang Lyu
2024-09-19 18:07:11 +08:00
committed by GitHub
parent 28f1353324 f65eca6723
commit 95051e5761
2 changed files with 11 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
cosyvoice/cli/model.py
View File

@@ -40,6 +40,8 @@ class CosyVoiceModel:
# hift cache
self.mel_cache_len = 20
self.source_cache_len = int(self.mel_cache_len * 256)
# speech fade in out
self.speech_window = np.hamming(2 * self.source_cache_len)
# rtf and decoding related
self.stream_scale_factor = 1
assert self.stream_scale_factor >= 1, 'stream_scale_factor should be greater than 1, change it according to your actual rtf'
@@ -114,13 +116,19 @@ class CosyVoiceModel:
self.mel_overlap_dict[uuid] = tts_mel[:, :, -self.mel_overlap_len:]
tts_mel = tts_mel[:, :, :-self.mel_overlap_len]
tts_speech, tts_source = self.hift.inference(mel=tts_mel, cache_source=hift_cache_source)
self.hift_cache_dict[uuid] = {'source': tts_source[:, :, -self.source_cache_len:], 'mel': tts_mel[:, :, -self.mel_cache_len:]}
if self.hift_cache_dict[uuid] is not None:
tts_speech = fade_in_out(tts_speech, self.hift_cache_dict[uuid]['speech'], self.speech_window)
self.hift_cache_dict[uuid] = {'mel': tts_mel[:, :, -self.mel_cache_len:],
'source': tts_source[:, :, -self.source_cache_len:],
'speech': tts_speech[:, -self.source_cache_len:]}
tts_speech = tts_speech[:, :-self.source_cache_len]
else:
if speed != 1.0:
assert self.hift_cache_dict[uuid] is None, 'speed change only support non-stream inference mode'
tts_mel = F.interpolate(tts_mel, size=int(tts_mel.shape[2] / speed), mode='linear')
tts_speech, tts_source = self.hift.inference(mel=tts_mel, cache_source=hift_cache_source)
if self.hift_cache_dict[uuid] is not None:
tts_speech = fade_in_out(tts_speech, self.hift_cache_dict[uuid]['speech'], self.speech_window)
return tts_speech
def inference(self, text, flow_embedding, llm_embedding=torch.zeros(0, 192),

4
cosyvoice/utils/common.py
View File

@@ -139,6 +139,6 @@ def fade_in_out(fade_in_mel, fade_out_mel, window):
device = fade_in_mel.device
fade_in_mel, fade_out_mel = fade_in_mel.cpu(), fade_out_mel.cpu()
mel_overlap_len = int(window.shape[0] / 2)
fade_in_mel[:, :, :mel_overlap_len] = fade_in_mel[:, :, :mel_overlap_len] * window[:mel_overlap_len] + \
fade_out_mel[:, :, -mel_overlap_len:] * window[mel_overlap_len:]
fade_in_mel[..., :mel_overlap_len] = fade_in_mel[..., :mel_overlap_len] * window[:mel_overlap_len] + \
fade_out_mel[..., -mel_overlap_len:] * window[mel_overlap_len:]
return fade_in_mel.to(device)