voxora-vad

Energy-based voice activity detection (VAD) for the voxora workspace. Tiny, deterministic, CPU-only, zero non-workspace runtime dependencies.

The detector slides a window over a stream of mono PCM samples, computes the RMS energy in each frame, and emits a VadSegment each time the speech/silence state machine transitions past the configured debounce windows.

Use cases

Trimming silence before a batch ASR pass. Pass the audio through [EnergyVad] first, then send only the speech runs to the engine.
Live-transcription UIs. Poll VadSegmenter::next_segment on each chunk of microphone PCM and only wake the decoder when is_speech == true.
Pre-filtering before speaker diarization. The diarizer only sees the speech runs, not the gaps.

Quick start

use voxora_vad::{EnergyVad, VadSegmenter, VadSegment};

let mut vad = EnergyVad::new();

let samples: Vec<f32> = vec![0.0; 16_000];
let seg: Option<VadSegment> = vad.next_segment(&samples);
assert!(seg.is_none()); // pure silence → no transition

let trailing: Option<VadSegment> = vad.flush();

The example in examples/segment_silence.rs runs the canonical voxora-testkit fixtures through the detector and prints the boundaries:

$ cargo run --example segment_silence
SILENCE_1S:
  segments: 0
sine_440hz_500ms:
  segments: 1
    [0, 8000) is_speech=true

Streaming integration

The voxora streaming engine trait (voxora_traits::StreamingAsrEngine) is intentionally not wired up by this crate. Streaming engines are still gated upstream on whisper-rs / candle incremental-decoding APIs (issues #50/#51). When the first streaming engine lands, a future patch will add a StreamingVadSegmenter extension that drives the decoder incrementally.

Crate conventions

Zero non-workspace runtime deps. The detector is pure f32 arithmetic over a sliding window, so the crate stays hermetic on a fresh cargo build. Matches the "tiny pure-Rust" stance of voxora-config.
Deterministic, no #[ignore] tests. The crate has no model/audio fixtures, so every test runs unconditionally and the CI test job stays fast.
ASR-specific. Not a generic signal-processing library. VAD here is a building block for the voxora speech-recognition stack only.

License