Struct Denoiser 

pub struct Denoiser { /* private fields */ }

RNNoise noise suppression as a mic-chain stage. One instance per stream — the network is stateful across frames.

Implementations

impl Denoiser

pub fn new(mic_hz: u32) -> Self

A denoiser for a microphone stream at mic_hz (the same rate given to pump_processed).

pub fn vad_probability(&self) -> f32

Speech probability in [0, 1] from RNNoise’s VAD head — the same recurrent features that drive the suppression gains, read out as a per-block classifier. Updated every 10 ms block the denoiser processes (the maximum across the blocks consumed by the most recent process_frame call); 0.0 before the first full block.

This is a learned VAD: it responds to the statistical fingerprint of speech (pitch movement, formants, syllabic modulation), not to level. Measured on the module-docs benchmark it separates cleanly — speech medians 0.76–0.96 against noise medians ≈ 0.00 for pink and street-traffic scenes (horns, engines) even at 0 dB SNR. Use it as the decision path for turn-taking (poll after each pumped frame, add hysteresis: on above ≈ 0.6, off below ≈ 0.3 with a ~300 ms hangover) where an energy-threshold VAD would false-trigger on loud non-speech.

Two measured caveats: babble and competing talkers score as speech (they are speech — see the module docs on level gating), and loud sustained broadband white noise from stream start can hold the head high until its noise estimate converges, so pair the probability with hysteresis rather than acting on a single block.

Trait Implementations

impl InputAudioProcessor for Denoiser

fn process_frame(&mut self, frame: &mut Vec<i16>)

Process one PCM16 frame in place (the frame may change length).