Merge branch '2noise:main' into webui

ChatTTS/core.py

@@ -95,6 +95,9 @@ def _load(
             assert gpt_ckpt_path, 'gpt_ckpt_path should not be None'
             gpt.load_state_dict(torch.load(gpt_ckpt_path, map_location='cpu'))
             self.pretrain_models['gpt'] = gpt
+            spk_stat_path = os.path.join(os.path.dirname(gpt_ckpt_path), 'spk_stat.pt')
+            assert os.path.exists(spk_stat_path), f'Missing spk_stat.pt: {spk_stat_path}'
+            self.pretrain_models['spk_stat'] = torch.load(spk_stat_path).to(device)
             self.logger.log(logging.INFO, 'gpt loaded.')
 
         if decoder_config_path:
@@ -144,6 +147,11 @@ def infer(
         wav = [self.pretrain_models['vocos'].decode(i).cpu().numpy() for i in mel_spec]
 
         return wav
+
+    def sample_random_speaker(self, ):
 
+        dim = self.pretrain_models['gpt'].gpt.layers[0].mlp.gate_proj.in_features
+        std, mean = self.pretrain_models['spk_stat'].chunk(2)
+        return torch.randn(dim, device=std.device) * std + mean
 
 

README.md

@@ -126,3 +126,6 @@ In the current released model, the only token-level control units are [laugh], [
 - [fish-speech](https://github.com/fishaudio/fish-speech) reveals capability of GVQ as audio tokenizer for LLM modeling.
 - [vocos](https://github.com/gemelo-ai/vocos) which is used as a pretrained vocoder.
 
+---
+## Special Appreciation
+- [wlu-audio lab](https://audio.westlake.edu.cn/) for early algorithm experiments.

README_CN.md

@@ -128,3 +128,6 @@ audio_array_en = chat.infer(inputs_en, params_refine_text=params_refine_text)
 - [fish-speech](https://github.com/fishaudio/fish-speech)一个优秀的自回归TTS模型, 揭示了GVQ用于LLM任务的可能性.
 - [vocos](https://github.com/gemelo-ai/vocos)作为模型中的vocoder.
 
+---
+## 特别致谢
+- [wlu-audio lab](https://audio.westlake.edu.cn/)为我们提供了早期算法试验的支持.