update: 优化vad判断逻辑

main/xiaozhi-server/config.yaml

@@ -454,7 +454,7 @@ VAD:
     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
+    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，实际延迟为:min_silence_duration_ms // 32 * 32ms
 
 LLM:
   # 所有openai类型均可以修改超参，以AliLLM为例

main/xiaozhi-server/core/connection.py

@@ -117,8 +117,8 @@ 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
+        self.continuous_false_count = 0
 
         # 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

main/xiaozhi-server/core/handle/receiveAudioHandle.py

@@ -33,7 +33,7 @@ async def handleAudioMessage(conn, audio):
 
 async def resume_vad_detection(conn):
     # 等待2秒后恢复VAD检测
-    await asyncio.sleep(1)
+    await asyncio.sleep(2)
     conn.just_woken_up = False
 
 

main/xiaozhi-server/core/providers/asr/base.py

@@ -68,7 +68,7 @@ class ASRProviderBase(ABC):
             conn.asr_audio.clear()
             conn.reset_vad_states()
 
-            if len(asr_audio_task) > 15:
+            if len(asr_audio_task) > 15 or conn.client_listen_mode == "manual":
                 await self.handle_voice_stop(conn, asr_audio_task)
 
     # 处理语音停止

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

@@ -1,6 +1,6 @@
 import time
-import numpy as np
 import torch
+import numpy as np
 import opuslib_next
 from config.logger import setup_logging
 from core.providers.vad.base import VADProviderBase
@@ -26,15 +26,12 @@ class VADProvider(VADProviderBase):
         threshold_low = config.get("threshold_low", "0.2")
         min_silence_duration_ms = config.get("min_silence_duration_ms", "1000")
 
+        # 双阈值判断参数
         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 = (
+        # 静默阈值，连续多少帧低于最低值才判断为静音
-            int(min_silence_duration_ms) if min_silence_duration_ms else 1000
+        self.stop_false_count = int(min_silence_duration_ms) // 32 if min_silence_duration_ms else 31
-        )
-
-        # 至少要多少帧才算有语音
-        self.frame_window_threshold = 3
 
     def is_vad(self, conn, opus_packet):
         try:
@@ -59,24 +56,24 @@ class VADProvider(VADProviderBase):
 
                 # 双阈值判断
                 if speech_prob >= self.vad_threshold:
-                    is_voice = True
+                    client_have_voice = True
                 elif speech_prob <= self.vad_threshold_low:
-                    is_voice = False
+                    client_have_voice = False
                 else:
-                    is_voice = conn.last_is_voice
+                    client_have_voice = conn.last_is_voice
+
+                if not client_have_voice:
+                    conn.continuous_false_count += 1
+                else:
+                    conn.continuous_false_count = 0
 
                 # 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
+                conn.last_is_voice = client_have_voice
 
-                # 更新滑动窗口
+                # 如果之前有声音，且连续为静音帧，达到静默阈值要求帧数，则认为已经说完一句话
-                conn.client_voice_window.append(is_voice)
+                if conn.client_have_voice and conn.continuous_false_count >= self.stop_false_count:
-                client_have_voice = (conn.client_voice_window.count(True) >= self.frame_window_threshold)
+                    conn.client_voice_stop = True
 
-                # 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:
-                    stop_duration = time.time() * 1000 - conn.last_activity_time
-                    if stop_duration >= self.silence_threshold_ms:
-                        conn.client_voice_stop = True
                 if client_have_voice:
                     conn.client_have_voice = True
                     conn.last_activity_time = time.time() * 1000