Integrate `SileroVAD` into `SpeechMonitor` for optional voice activity detection. Add `_detect_speech()` method and update queue handling logic. Implement `SileroVAD` as a singleton for efficient model management.

src/voice_dialogue/services/audio/vad.py

@@ -0,0 +1,88 @@
+from typing import Optional
+
+import numpy as np
+import torch
+from silero_vad import load_silero_vad
+
+from voice_dialogue.utils.logger import logger
+
+
+class SileroVAD:
+    """
+    一个线程安全的、基于单例模式的Silero VAD模型包装器。
+
+    该类在首次实例化时加载 Silero VAD 模型，并提供一个方法来检测音频帧中的语音活动。
+    设计为单例可以避免在应用中重复加载这个较为消耗资源模型。
+    """
+    _instance: Optional['SileroVAD'] = None
+    _model = None
+
+    def __new__(cls, *args, **kwargs):
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+        return cls._instance
+
+    def __init__(self, threshold: float = 0.7):
+        """
+        初始化 Silero VAD 模型。模型只会在首次创建实例时加载。
+
+        Args:
+            threshold (float): 用于判定语音活动的置信度阈值 (范围 0.0 到 1.0)。
+        """
+        if self._model is None:
+            logger.info("正在首次初始化 Silero VAD 模型...")
+            try:
+                self._model = load_silero_vad()
+                self._model.reset_states()
+                self.threshold = threshold
+                logger.info("Silero VAD 模型初始化成功。")
+            except Exception as e:
+                logger.error(f"初始化 Silero VAD 模型失败: {e}", exc_info=True)
+                # 如果失败，重置实例，以便下次可以重试
+                SileroVAD._instance = None
+                raise
+
+    def is_voice_active(self, audio_frame: np.ndarray, sample_rate: int = 16000) -> bool:
+        """
+        检测给定的音频帧中是否包含语音活动。
+
+        Args:
+            audio_frame (np.ndarray): 一个一维的 float32 numpy 数组，代表音频数据。
+                                      其数值范围应为 [-1.0, 1.0]。
+                                      对于16kHz采样率，帧大小必须为 [512, 1024, 1536] 之一。
+            sample_rate (int): 音频的采样率，必须是 8000 或 16000。
+
+        Returns:
+            bool: 如果检测到语音活动，返回 True，否则返回 False。
+        """
+        if self._model is None:
+            logger.error("VAD 模型未初始化，无法执行检测。")
+            return False
+
+        if not isinstance(audio_frame, np.ndarray):
+            logger.warning("VAD 检测的输入必须是一个 numpy 数组。")
+            return False
+
+        # Silero VAD 模型要求 float32 类型
+        if audio_frame.dtype != np.float32:
+            audio_frame = audio_frame.astype(np.float32)
+
+        window_size = 512 if sample_rate == 16000 else 256
+
+        audio_tensor = torch.from_numpy(audio_frame)
+
+        try:
+            probs = []
+            for i in range(0, len(audio_tensor), window_size):
+                audio_slice = audio_tensor[i:i + window_size]
+                if len(audio_slice) < window_size:
+                    audio_slice = audio_tensor[-window_size:]
+
+                # 模型会返回一个包含语音可能性的张量
+                prob = self._model(audio_slice, sample_rate).item()
+                probs.append(prob)
+
+            return np.max(probs) > self.threshold
+        except Exception as e:
+            logger.error(f"VAD 检测过程中发生错误: {e}")
+            return False

src/voice_dialogue/services/speech/monitor.py

@@ -19,6 +19,7 @@ from voice_dialogue.core.constants import (
 )
 from voice_dialogue.core.enums import AudioState
 from voice_dialogue.models.voice_task import VoiceTask
+from voice_dialogue.services.audio.vad import SileroVAD
 from voice_dialogue.utils.logger import logger
 
 
@@ -49,6 +50,7 @@ class SpeechStateMonitor(BaseThread):
             self, group=None, target=None, name=None, args=(), kwargs=None, *, daemon=None,
             audio_frame_queue: Queue,
             user_voice_queue: Queue,
+            enable_vad: bool = False,
     ):
         """
         初始化语音状态监控器
@@ -56,12 +58,18 @@ class SpeechStateMonitor(BaseThread):
         Args:
             audio_frame_queue: 音频帧队列
             user_voice_queue: 用户语音队列
+            enable_vad: 是否启用语音活动检测
         """
         super().__init__(group, target, name, args, kwargs, daemon=daemon)
 
         self.audio_frame_queue = audio_frame_queue
         self.user_voice_queue = user_voice_queue
         self.sample_rate = 16000
+        self._enable_vad = enable_vad
+
+        self._vad_instance = None
+        if self._enable_vad:
+            self._vad_instance = SileroVAD()
 
         # 配置参数
         self.config = SpeechMonitorConfig()
@@ -104,11 +112,19 @@ class SpeechStateMonitor(BaseThread):
         """将 int16 格式的音频字节数据转换为 [-1.0, 1.0] 范围的 numpy 浮点数组。"""
         return np.frombuffer(data, dtype=np.int16).astype(np.float32) / np.iinfo(np.int16).max
 
+    def _detect_speech(self, audio_frame: np.ndarray) -> bool:
+        return self._vad_instance.is_voice_active(audio_frame, self.sample_rate)
+
     def _get_audio_frame_from_queue(self):
         """从队列获取音频帧"""
         try:
-            data, is_voice_active = self.audio_frame_queue.get(block=False, timeout=self.config.QUEUE_TIMEOUT)
-            audio_frame = self._normalize_audio_frame(data)
+            if self._enable_vad:
+                data = self.audio_frame_queue.get(block=False, timeout=self.config.QUEUE_TIMEOUT)
+                audio_frame = self._normalize_audio_frame(data)
+                is_voice_active = self._detect_speech(audio_frame)
+            else:
+                data, is_voice_active = self.audio_frame_queue.get(block=False, timeout=self.config.QUEUE_TIMEOUT)
+                audio_frame = self._normalize_audio_frame(data)
             return audio_frame, is_voice_active
         except Empty:
             return None, None