Expand description
Segment a mono signal into voiced regions with hysteresis and hangover.
A raw per-frame decision (crate::features::vad::frame_is_voiced) chatters
at speech onsets/offsets and on brief dropouts (stops, glottal gaps). This
module turns the frame-level proposal stream into stable [start, end)
frame-index segments using two duration guards:
- min-speech (onset debounce): a run of voiced frames must reach a minimum length before a segment opens, rejecting isolated spurious frames.
- min-silence (hangover): once open, a segment only closes after a sustained run of unvoiced frames, so a short dropout inside speech does not split it.
The noise floor is adaptive and streamed online: a fixed stack ring of
the most recent per-frame energies feeds
crate::features::vad::noise_floor_min_stat, so the floor tracks the quiet
frames without any heap allocation.
EPISTEMIC RULE: output segments are proposals. Frames at or below the adaptive floor stay unvoiced — silence is never coerced into a segment.
Functions§
- segment_
voiced - Detect voiced segments in
signal, writing[start_frame, end_frame)index pairs intoout_segments.