refactor: replace Kokoro-82M with suno/bark for TTS, update HF cache path, and add model warmup on startup.

backend/tts_asr.py

@@ -283,8 +283,9 @@ def _check_model_cached(model_id: str) -> bool:
         return True  # 非离线模式，不检查缓存
     
     try:
-        from transformers import file_utils
-        cache_dir = file_utils.default_cache_path
+        import os
+        from huggingface_hub.constants import HF_HUB_CACHE
+        cache_dir = HF_HUB_CACHE
         
         # 简单的缓存检查：查找模型目录
         model_name = model_id.replace("/", "--")
@@ -350,14 +351,14 @@ async def _load_tts_pipeline_with_retry(max_retries: int = 2) -> bool:
                     device_to_use = _device_arg()
                     torch_dtype = _get_torch_dtype()
                     
-                    model_id = "hexgrad/Kokoro-82M"
+                    model_id = "suno/bark"
                     
                     # 离线模式检查
                     if TTS_ASR_OFFLINE_MODE and not _check_model_cached(model_id):
                         logger.error("[TTS] 离线模式下模型 %s 未缓存", model_id)
                         return False
 
-                    logger.info("[TTS] 加载 Kokoro-82M 模型 (尝试 %d/%d, 设备: %s)...", 
+                    logger.info("[TTS] 加载 suno/bark 模型 (尝试 %d/%d, 设备: %s)...", 
                                attempt + 1, max_retries, device_to_use)
 
                     _tts_pipeline = await asyncio.to_thread(
@@ -370,7 +371,7 @@ async def _load_tts_pipeline_with_retry(max_retries: int = 2) -> bool:
                         )
                     )
 
-                    logger.info("[TTS] Kokoro-82M 模型加载完成")
+                    logger.info("[TTS] suno/bark 模型加载完成")
                     return True
 
                 except RuntimeError as e:
@@ -552,7 +553,7 @@ async def _warmup_tts() -> bool:
 
         def warmup_inference():
             try:
-                result = tts(TTS_WARMUP_TEXT, voice="af_bella")
+                result = tts(TTS_WARMUP_TEXT)
                 if isinstance(result, dict):
                     audio = result.get("audio")
                     if hasattr(audio, "cpu"):
@@ -759,7 +760,7 @@ async def _tts_sync_with_retry(text: str, voice: str = "af_bella", rate: float =
     for attempt in range(max_retries):
         try:
             def inference():
-                result = tts(text, voice=voice)
+                result = tts(text)
                 audio = None
                 sr = sample_rate
 
@@ -770,7 +771,7 @@ async def _tts_sync_with_retry(text: str, voice: str = "af_bella", rate: float =
                     audio = result[0]
 
                 if audio is None:
-                    raise RuntimeError("Kokoro 未返回音频数据")
+                    raise RuntimeError("TTS模型未返回音频数据")
 
                 if hasattr(audio, "cpu"):
                     audio = audio.cpu().numpy()
@@ -1017,7 +1018,7 @@ async def get_config(api_key: str = Security(get_api_key)):
             "cuda_memory_limit_mb": caps.cuda_memory_limit_mb,
         },
         "model": {
-            "tts": "hexgrad/Kokoro-82M",
+            "tts": "suno/bark",
             "asr_current_size": _asr_model_size,
             "asr_recommended_size": caps.recommended_model_size,
             "available_sizes": list(WHISPER_MODEL_SIZES.keys()),