fix: run TTS synthesis in thread pool to avoid blocking event loop

Also add size check for int8 model to skip LFS pointer files.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

backend/main/xiaozhi-server/core/providers/tts/sherpa_tts.py

@@ -1,5 +1,7 @@
 import io
+import os
 import wave
+import asyncio
 import numpy as np
 import sherpa_onnx
 from config.logger import setup_logging
@@ -19,8 +21,7 @@ class TTSProvider(TTSProviderBase):
 
         # 优先使用 int8 量化模型（更快）
         model_file = f"{model_dir}/model.int8.onnx"
-        import os
-        if not os.path.exists(model_file):
+        if not os.path.exists(model_file) or os.path.getsize(model_file) < 1024:
             model_file = f"{model_dir}/model.onnx"
 
         num_threads = int(config.get("num_threads", 8))
@@ -44,7 +45,8 @@ class TTSProvider(TTSProviderBase):
             f"SherpaOnnxTTS 初始化完成: model_dir={model_dir}, sample_rate={self.sample_rate}, sid={self.sid}"
         )
 
-    async def text_to_speak(self, text, output_file):
+    def _generate_wav(self, text):
+        """同步合成，在线程池中调用"""
         audio = self.tts.generate(text, sid=self.sid, speed=self.speed)
         samples = np.array(audio.samples, dtype=np.float32)
         pcm = (samples * 32767).astype(np.int16)
@@ -55,7 +57,11 @@ class TTSProvider(TTSProviderBase):
             wf.setsampwidth(2)
             wf.setframerate(self.sample_rate)
             wf.writeframes(pcm.tobytes())
-        wav_data = wav_io.getvalue()
+        return wav_io.getvalue()
+
+    async def text_to_speak(self, text, output_file):
+        loop = asyncio.get_event_loop()
+        wav_data = await loop.run_in_executor(None, self._generate_wav, text)
 
         if output_file:
             with open(output_file, "wb") as f: