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

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/cli/cosyvoice.py
View File

@@ -74,6 +74,9 @@ class CosyVoice:
self.frontend.spk2info[zero_shot_spk_id] = model_input
return True
def save_spkinfo(self):
torch.save(self.frontend.spk2info, '{}/spk2info.pt'.format(self.model_dir))
def inference_sft(self, tts_text, spk_id, stream=False, speed=1.0, text_frontend=True):
for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
model_input = self.frontend.frontend_sft(i, spk_id)
@@ -99,9 +102,9 @@ class CosyVoice:
yield model_output
start_time = time.time()
def inference_cross_lingual(self, tts_text, prompt_speech_16k, stream=False, speed=1.0, text_frontend=True):
def inference_cross_lingual(self, tts_text, prompt_speech_16k, zero_shot_spk_id='', stream=False, speed=1.0, text_frontend=True):
for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
model_input = self.frontend.frontend_cross_lingual(i, prompt_speech_16k, self.sample_rate)
model_input = self.frontend.frontend_cross_lingual(i, prompt_speech_16k, self.sample_rate, zero_shot_spk_id)
start_time = time.time()
logging.info('synthesis text {}'.format(i))
for model_output in self.model.tts(**model_input, stream=stream, speed=speed):
@@ -174,10 +177,10 @@ class CosyVoice2(CosyVoice):
def inference_instruct(self, *args, **kwargs):
raise NotImplementedError('inference_instruct is not implemented for CosyVoice2!')
def inference_instruct2(self, tts_text, instruct_text, prompt_speech_16k, stream=False, speed=1.0, text_frontend=True):
def inference_instruct2(self, tts_text, instruct_text, prompt_speech_16k, zero_shot_spk_id='', stream=False, speed=1.0, text_frontend=True):
assert isinstance(self.model, CosyVoice2Model), 'inference_instruct2 is only implemented for CosyVoice2!'
for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
model_input = self.frontend.frontend_instruct2(i, instruct_text, prompt_speech_16k, self.sample_rate)
model_input = self.frontend.frontend_instruct2(i, instruct_text, prompt_speech_16k, self.sample_rate, zero_shot_spk_id)
start_time = time.time()
logging.info('synthesis text {}'.format(i))
for model_output in self.model.tts(**model_input, stream=stream, speed=speed):