Expand description
Adaptive noise-floor estimation by minimum statistics.
The quietest recent frames are, with high probability, noise-only. Tracking the minimum short-term energy over a trailing window therefore yields a robust estimate of the noise level that follows slow drifts (fans, hiss, room tone) without ever being pulled up by transient speech or tones.
The raw minimum systematically under-estimates the mean noise power (it is an extreme-order statistic), so it is multiplied by a fixed bias-compensation factor to recover an estimate of the floor an activity gate should clear.
Zero-heap: a single pass over a caller-supplied energy slice. Stateless, so
callers stream by keeping their own small ring of recent per-frame energies
(see crate::features::vad::segment_voiced).
Constants§
- NOISE_
BIAS - Bias compensation for the minimum-of-window estimator.
Functions§
- noise_
floor_ min_ stat - Estimate the noise floor as the bias-compensated minimum energy over the
trailing
windowentries ofenergies(per-frame RMS or power values).