T

app.py

@@ -256,7 +256,6 @@ def infer_voice_clone( part, language,audio_tuple,ref_text,use_xvector_only):
 @spaces.GPU
 def infer_voice_clone_from_prompt(part, language, prompt_file_path):
     """Single segment inference for Voice Clone using pre-extracted prompt."""
-    seed_everything(42)
     logger.info("正在加载音频特征文件...")
     voice_clone_prompt = torch.load(prompt_file_path, map_location='cuda', weights_only=False)
     logger.info("音频特征文件加载成功。")
@@ -268,9 +267,10 @@ def infer_voice_clone_from_prompt(part, language, prompt_file_path):
         voice_clone_prompt=voice_clone_prompt,
         max_new_tokens=2048,
         # 核心参数：固定 seed
-        seed=42, 
-        temperature=0.3, # 配合低温度，音色会更稳
-        top_p=0.85
+        seed=4222, 
+        temperature=0.45, # 配合低温度，音色会更稳
+        top_p=0.9,
+        repetition_penalty=1.2 # 惩罚重复，防止停顿或复读机现象
     )
     return wavs[0], sr