VAD

Entry point group: opm.VAD Config entry point group: opm.VAD.config Template: ovos_plugin_manager.templates.vad

---

VADEngine base class

from ovos_plugin_manager.templates.vad import VADEngine

VAD plugins determine whether a given audio segment contains speech. They are used by ovos-dinkum-listener to trim leading/trailing silence before STT.

Constructor

VADEngine(config: Optional[dict] = None, sample_rate: Optional[int] = None)

Attribute	Type	Default	Description
sample_rate	int	16000	Audio sample rate in Hz.
padding_duration_ms	int	300	Ring buffer duration in ms.
frame_duration_ms	int	30	Frame size in ms.
thresh	float	0.8	Fraction of voiced frames needed to trigger speech detection.
num_padding_frames	int	derived	padding_duration_ms / frame_duration_ms.

Abstract method (must implement)

is_silence(chunk: bytes) -> bool

Return True if the provided audio chunk does not contain speech.

Provided methods

extract_speech(audio: bytes) -> bytes

Remove leading and trailing silence from audio using a sliding ring-buffer algorithm. Returns only the voiced portion of the audio.

reset()

Reset any internal state. Called between utterances.

runtime_requirements (classproperty)

Defaults to fully offline.

---

AudioFrame

from ovos_plugin_manager.templates.vad import AudioFrame

Data class representing a single frame of audio.

Attribute	Type	Description
bytes	bytes	Raw PCM audio data.
timestamp	float	Start time of the frame in seconds.
duration	float	Duration of the frame in seconds.

---

Configuration

VAD config is read from ovos.conf under listener.VAD:

{
  "listener": {
    "VAD": {
      "module": "ovos-vad-plugin-silero",
      "ovos-vad-plugin-silero": {
        "thresh": 0.5,
        "padding_duration_ms": 400
      }
    }
  }
}