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

fix short tts_text bug

Browse Source
This commit is contained in:
lyuxiang.lx
2024-10-22 12:56:29 +08:00
parent 0930b4a106
commit 16d66dc6a6
2 changed files with 5 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
cosyvoice/cli/cosyvoice.py
View File

@@ -67,6 +67,8 @@ class CosyVoice:
def inference_zero_shot(self, tts_text, prompt_text, prompt_speech_16k, stream=False, speed=1.0): def inference_zero_shot(self, tts_text, prompt_text, prompt_speech_16k, stream=False, speed=1.0):
prompt_text = self.frontend.text_normalize(prompt_text, split=False) prompt_text = self.frontend.text_normalize(prompt_text, split=False)
for i in tqdm(self.frontend.text_normalize(tts_text, split=True)): for i in tqdm(self.frontend.text_normalize(tts_text, split=True)):
if len(i) < 0.5 * len(prompt_text):
logging.warning('synthesis text {} too short than prompt text {}, this may lead to bad performance'.format(i, prompt_text))
model_input = self.frontend.frontend_zero_shot(i, prompt_text, prompt_speech_16k) model_input = self.frontend.frontend_zero_shot(i, prompt_text, prompt_speech_16k)
start_time = time.time() start_time = time.time()
logging.info('synthesis text {}'.format(i)) logging.info('synthesis text {}'.format(i))

3
cosyvoice/llm/llm.py
View File

@@ -202,6 +202,9 @@ class TransformerLM(torch.nn.Module):
att_mask=torch.tril(torch.ones((1, lm_input.shape[1], lm_input.shape[1]), att_mask=torch.tril(torch.ones((1, lm_input.shape[1], lm_input.shape[1]),
device=lm_input.device)).to(torch.bool)) device=lm_input.device)).to(torch.bool))
logp = self.llm_decoder(y_pred[:, -1]).log_softmax(dim=-1) logp = self.llm_decoder(y_pred[:, -1]).log_softmax(dim=-1)
# force continue decode first token
if i == 0:
logp[:, self.speech_token_size] = -float('inf')
top_ids = self.sampling_ids(logp.squeeze(dim=0), out_tokens, sampling, ignore_eos=True if i < min_len else False).item() top_ids = self.sampling_ids(logp.squeeze(dim=0), out_tokens, sampling, ignore_eos=True if i < min_len else False).item()
if top_ids == self.speech_token_size: if top_ids == self.speech_token_size:
break break