Загрузка исходника

wyoming_satellite/vad.py

@@ -0,0 +1,88 @@
+"""Voice activity detection."""
+from typing import Optional
+import math
+import array
+
+
+class SileroVad:
+    """Voice activity detection with silero VAD."""
+
+    def __init__(self, threshold: float, trigger_level: int) -> None:
+        from pysilero_vad import SileroVoiceActivityDetector
+
+        self.detector = SileroVoiceActivityDetector()
+        self.threshold = threshold
+        self.trigger_level = trigger_level
+        self._activation = 0
+
+    def __call__(self, audio_bytes: Optional[bytes]) -> bool:
+        if audio_bytes is None:
+            # Reset
+            self._activation = 0
+            self.detector.reset()
+            return False
+
+        if self.detector(audio_bytes) >= self.threshold:
+            # Speech detected
+            self._activation += 1
+            if self._activation >= self.trigger_level:
+                self._activation = 0
+                return True
+        else:
+            # Silence detected
+            self._activation = max(0, self._activation - 1)
+
+        return False
+
+
+class WebRtcVadEnergy:
+    """Voice activity detection using webrtcvad + RMS energy threshold.
+
+    Expects 16-bit mono PCM at 16Khz. Chunks are internally processed in 20ms windows.
+    """
+
+    def __init__(self, trigger_level: int, energy_threshold: Optional[float] = None) -> None:
+        import webrtcvad  # type: ignore
+
+        self.vad = webrtcvad.Vad(3)
+        self.trigger_level = trigger_level
+        self.energy_threshold = energy_threshold if energy_threshold is not None else 200.0
+        self._activation = 0
+
+        # 20ms @16kHz, 16-bit mono
+        self._samples_per_window = 320
+        self._bytes_per_window = self._samples_per_window * 2
+
+    def __call__(self, audio_bytes: Optional[bytes]) -> bool:
+        if audio_bytes is None:
+            # Reset
+            self._activation = 0
+            return False
+
+        # Process in 20ms windows as required by webrtcvad
+        for i in range(0, len(audio_bytes) - (len(audio_bytes) % self._bytes_per_window), self._bytes_per_window):
+            window = audio_bytes[i : i + self._bytes_per_window]
+
+            # RMS energy
+            pcm = array.array("h")
+            pcm.frombytes(window)
+            if len(pcm) == 0:
+                continue
+            rms = math.sqrt(sum(int(s) * int(s) for s in pcm) / len(pcm))
+
+            is_speech = False
+            try:
+                is_speech = self.vad.is_speech(window, 16000) and (rms >= float(self.energy_threshold))
+            except Exception:
+                # If VAD errors on malformed window, skip
+                is_speech = False
+
+            if is_speech:
+                self._activation += 1
+                if self._activation >= self.trigger_level:
+                    self._activation = 0
+                    return True
+            else:
+                self._activation = max(0, self._activation - 1)
+
+        return False