update: TTS复用链接，VAD双阈值判断

main/xiaozhi-server/config.yaml

@@ -325,6 +325,7 @@ VAD:
   SileroVAD:
     type: silero
     threshold: 0.5
+    threshold_low: 0.3
     model_dir: models/snakers4_silero-vad
     min_silence_duration_ms: 200  # 如果说话停顿比较长，可以把这个值设置大一些
 

main/xiaozhi-server/core/connection.py

@@ -17,6 +17,7 @@ from core.utils.util import (
     filter_sensitive_info,
 )
 from typing import Dict, Any
+from collections import deque
 from core.utils.modules_initialize import (
     initialize_modules,
     initialize_tts,
@@ -112,6 +113,8 @@ class ConnectionHandler:
         self.client_have_voice = False
         self.last_activity_time = 0.0  # 统一的活动时间戳（毫秒）
         self.client_voice_stop = False
+        self.client_voice_window = deque(maxlen=5)
+        self.last_is_voice = False
 
         # asr相关变量
         # 因为实际部署时可能会用到公共的本地ASR，不能把变量暴露给公共ASR
@@ -858,6 +861,9 @@ class ConnectionHandler:
             elif self.websocket:
                 await self.websocket.close()
 
+            if self.tts:
+                await self.tts.close()
+
             # 最后关闭线程池（避免阻塞）
             if self.executor:
                 self.executor.shutdown(wait=False)

main/xiaozhi-server/core/providers/tts/huoshan_double_stream.py

@@ -1,4 +1,3 @@
-import os
 import uuid
 import json
 import queue
@@ -10,7 +9,6 @@ from config.logger import setup_logging
 from core.utils import opus_encoder_utils
 from core.utils.util import check_model_key
 from core.providers.tts.base import TTSProviderBase
-from core.handle.abortHandle import handleAbortMessage
 from core.providers.tts.dto.dto import SentenceType, ContentType, InterfaceType
 from asyncio import Task
 
@@ -175,6 +173,9 @@ class TTSProvider(TTSProviderBase):
     async def _ensure_connection(self):
         """建立新的WebSocket连接"""
         try:
+            if self.ws:
+                logger.bind(tag=TAG).info(f"使用已有链接...")
+                return self.ws
             logger.bind(tag=TAG).info("开始建立新连接...")
             ws_header = {
                 "X-Api-App-Key": self.appId,
@@ -368,9 +369,17 @@ class TTSProvider(TTSProviderBase):
                 logger.bind(tag=TAG).info("会话结束请求已发送")
 
                 # 等待监听任务完成
-                if hasattr(self, "_monitor_task"):
+                if hasattr(self, "_monitor_task") and self._monitor_task is not None:
                     try:
-                        await self._monitor_task
+                        # 设置4秒超时等待监听任务结束
+                        await asyncio.wait_for(self._monitor_task, timeout=4)
+                    except asyncio.TimeoutError:
+                        logger.bind(tag=TAG).warning("等待监听任务超时，强制取消")
+                        self._monitor_task.cancel()
+                        try:
+                            await self._monitor_task
+                        except asyncio.CancelledError:
+                            pass
                     except Exception as e:
                         logger.bind(tag=TAG).error(
                             f"等待监听任务完成时发生错误: {str(e)}"
@@ -378,8 +387,6 @@ class TTSProvider(TTSProviderBase):
                     finally:
                         self._monitor_task = None
 
-                # 关闭连接
-                await self.close()
         except Exception as e:
             logger.bind(tag=TAG).error(f"关闭会话失败: {str(e)}")
             # 确保清理资源
@@ -400,6 +407,15 @@ class TTSProvider(TTSProviderBase):
 
     async def close(self):
         """资源清理方法"""
+        # 取消监听任务
+        if self._monitor_task:
+            try:
+                self._monitor_task.cancel()
+                await self._monitor_task
+            except asyncio.CancelledError:
+                pass
+            self._monitor_task = None
+
         if self.ws:
             try:
                 await self.ws.close()
@@ -413,6 +429,7 @@ class TTSProvider(TTSProviderBase):
         is_first_sentence = True
         first_sentence_segment_count = 0  # 添加计数器
         try:
+            session_finished = False  # 标记会话是否正常结束
             while not self.conn.stop_event.is_set():
                 try:
                     # 确保 `recv()` 运行在同一个 event loop
@@ -466,6 +483,7 @@ class TTSProvider(TTSProviderBase):
                     elif res.optional.event == EVENT_SessionFinished:
                         logger.bind(tag=TAG).debug(f"会话结束～～")
                         self._process_before_stop_play_files()
+                        session_finished = True
                         break
                 except websockets.ConnectionClosed:
                     logger.bind(tag=TAG).warning("WebSocket连接已关闭")
@@ -476,15 +494,15 @@ class TTSProvider(TTSProviderBase):
                     )
                     traceback.print_exc()
                     break
-        finally:
-            # 确保清理资源
-            if self.ws:
+            # 仅在连接异常时才关闭
+            if not session_finished and self.ws:
                 try:
                     await self.ws.close()
                 except:
                     pass
                 self.ws = None
             # 监听任务退出时清理引用
+        finally:
             self._monitor_task = None
 
     async def send_event(

main/xiaozhi-server/core/providers/vad/silero.py

@@ -23,22 +23,24 @@ class VADProvider(VADProviderBase):
 
         # 处理空字符串的情况
         threshold = config.get("threshold", "0.5")
+        threshold_low = config.get("threshold_low", "0.2")
         min_silence_duration_ms = config.get("min_silence_duration_ms", "1000")
 
         self.vad_threshold = float(threshold) if threshold else 0.5
+        self.vad_threshold_low = float(threshold_low) if threshold_low else 0.2
+
         self.silence_threshold_ms = (
             int(min_silence_duration_ms) if min_silence_duration_ms else 1000
         )
 
+        # 至少要多少帧才算有语音
+        self.frame_window_threshold = 3
+
     def is_vad(self, conn, opus_packet):
         try:
             pcm_frame = self.decoder.decode(opus_packet, 960)
             conn.client_audio_buffer.extend(pcm_frame)  # 将新数据加入缓冲区
 
-            # 确保帧计数器存在
-            if not hasattr(conn, "client_voice_frame_count"):
-                conn.client_voice_frame_count = 0
-
             # 处理缓冲区中的完整帧（每次处理512采样点）
             client_have_voice = False
             while len(conn.client_audio_buffer) >= 512 * 2:
@@ -54,15 +56,21 @@ class VADProvider(VADProviderBase):
                 # 检测语音活动
                 with torch.no_grad():
                     speech_prob = self.model(audio_tensor, 16000).item()
-                is_voice = speech_prob >= self.vad_threshold
 
-                if is_voice:
-                    conn.client_voice_frame_count += 1
+                # 双阈值判断
+                if speech_prob >= self.vad_threshold:
+                    is_voice = True
+                elif speech_prob <= self.vad_threshold_low:
+                    is_voice = False
                 else:
-                    conn.client_voice_frame_count = 0
+                    is_voice = conn.last_is_voice
 
-                # 只有连续4帧检测到语音才认为有语音
-                client_have_voice = conn.client_voice_frame_count >= 4
+                # 声音没低于最低值则延续前一个状态，判断为有声音
+                conn.last_is_voice = is_voice
+
+                # 更新滑动窗口
+                conn.client_voice_window.append(is_voice)
+                client_have_voice = (conn.client_voice_window.count(True) >= self.frame_window_threshold)
 
                 # 如果之前有声音，但本次没有声音，且与上次有声音的时间差已经超过了静默阈值，则认为已经说完一句话
                 if conn.client_have_voice and not client_have_voice: