xr_ai_vad

Shared Silero-VAD utterance detector for xr-ai agent workers.

Consumes int16 LE PCM bytes and emits int16 PCM utterance bytes via an async callback when speech ends.

Attributes

OnSpeechStartCb	Async callback fired once when an utterance crosses the speech threshold.
OnUtteranceCb	Async callback receiving completed int16 PCM bytes and their sample rate.

Classes

VadDetector

Per-participant Silero VAD + utterance accumulator.

Package Contents

xr_ai_vad.OnSpeechStartCb

Async callback fired once when an utterance crosses the speech threshold.

xr_ai_vad.OnUtteranceCb

Async callback receiving completed int16 PCM bytes and their sample rate.

class xr_ai_vad.VadDetector(

on_utterance: OnUtteranceCb,

*,

on_speech_start: OnSpeechStartCb | None = None,

silence_duration: float = 0.8,

min_speech: float = 0.15,

silero_threshold: float = 0.5,

)

Per-participant Silero VAD + utterance accumulator.

feed() accepts raw int16 LE PCM bytes. When a complete utterance is detected (silence after min_speech) the on_utterance callback is awaited with int16 PCM bytes and the sample rate.

When on_speech_start is provided, it fires once per utterance at the moment speech_s first crosses min_speech — a “leading edge” hook for speculative work (e.g. warming up a downstream resource before STT completes). It is awaited on the feeding task before processing further windows or finalizing that utterance. Callback exceptions are logged and isolated so detection can continue.

reset() → None

Drop any in-progress utterance without emitting it.

async feed(pcm_int16: bytes, sample_rate: int) → None

Process one chunk of int16 LE PCM audio.

pcm_int16 — raw int16 little-endian PCM bytes (mono). sample_rate — sample rate in Hz (constant for this stream).

Incomplete classification windows remain buffered. Speech and silence durations advance only when a complete 32 ms window is classified, independently of the caller’s chunk size.