Merge pull request #618 from FunAudioLLM/dev/lyuxiang.lx

Dev/lyuxiang.lx

Author: aluminumbox

cosyvoice/cli/cosyvoice.py

@@ -67,6 +67,8 @@ def inference_sft(self, tts_text, spk_id, 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)
         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)
             start_time = time.time()
             logging.info('synthesis text {}'.format(i))

cosyvoice/llm/llm.py

@@ -202,6 +202,9 @@ def inference(
                                                                   att_mask=torch.tril(torch.ones((1, lm_input.shape[1], lm_input.shape[1]),
                                                                                                  device=lm_input.device)).to(torch.bool))
             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()
             if top_ids == self.speech_token_size:
                 break

examples/libritts/cosyvoice/conf/cosyvoice.fromscratch.yaml

@@ -141,7 +141,7 @@ mel_spec_transform1: !name:matcha.utils.audio.mel_spectrogram
     hop_size: 256
     win_size: 1024
     fmin: 0
-    fmax: 8000
+    fmax: null
     center: False
 hifigan: !new:cosyvoice.hifigan.hifigan.HiFiGan
     generator: !ref <hift>

examples/libritts/cosyvoice/conf/cosyvoice.yaml

@@ -141,7 +141,7 @@ mel_spec_transform1: !name:matcha.utils.audio.mel_spectrogram
     hop_size: 256
     win_size: 1024
     fmin: 0
-    fmax: 8000
+    fmax: null
     center: False
 hifigan: !new:cosyvoice.hifigan.hifigan.HiFiGan
     generator: !ref <hift>

requirements.txt

@@ -6,6 +6,7 @@ gdown==5.1.0
 gradio==4.32.2
 grpcio==1.57.0
 grpcio-tools==1.57.0
+huggingface-hub==0.23.5
 hydra-core==1.3.2
 HyperPyYAML==1.2.2
 inflect==7.3.1