@audio/denoise-deplosive

import deplosive from '@audio/denoise-deplosive'

Takes a close-mic p or b out: a pressure pulse that rises out of nothing under 80 Hz and has no period, where a voice's or a bass note's low end repeats at its pitch. A pop begins when the 3 ms envelope of the band under 80 Hz jumps over 3× its 30 ms average while standing over triggerRatio× the band over 120 Hz (a voice's fundamental and up), and holds while that lasts and the low band stays aperiodic: its normalized autocorrelation's peak at a 2.5–25 ms lag (Talkin 1995) under ½, where a harmonic part would lead (Boersma 1993). The voice band is floored 6 dB under its peak over the last seconds: a room's rumble in a pause is no pop, and a pop reaches the voice's level. Removal takes the sound's own part under crossover away, x − w·LP(x), LP linear-phase: a pulse through a causal filter leaves the filter's response to its edges behind (with the pops' spans known, a duck to a causal 2nd-order high-pass at 100 Hz takes a median 12 dB of their error, the linear-phase band at 120 Hz 21). LP runs at 2 kHz (three box sums decimate, a Kaiser-windowed sinc low-passes, linear interpolation returns), and its half-length is the look-ahead: the output is latency(fs) samples late (14 ms, declared to the host), the duck starts 4 ms before where the pop's low band rose and is held to its end. With no pop the output is the input, sample for sample, that late.

deplosive(data, { fs: 48000 })                  // in place, aligned
let write = deplosive({ fs: 48000 })            // stream: each block back whole, latency(48000) samples late; write() flushes

Against iZotope RX 12 De-plosive, node bench/rx/deplosive.mjs in audio (2026-10): VoiceBank clean speech (2 test speakers, 11.4 min), pops added at stop bursts found on it (Liu 1996's +b landmark), six in ten: a one-sided pressure pulse of 20–80 ms rising in 1–4 ms (as the narrations' own pops do), under 200 Hz, 0.3–2× the speech peak. Every setting chosen on 28 training speakers; RX tuned: sensitivity 1, strength 5, frequency limit 125 Hz (defaults 5, 5, 200). The error to the clean speech taken away over each pop and the 100 ms after it; SNR to the clean speech over the reels with pops; on the speech alone, its voiced frames:

0.2.0 missed one pop in ten (a pop under a word's level never stood 4× over the band above 200 Hz) and took 6 dB of the rest. node scripts/lowend.js deplosive puts clean speech and music through it, then speech with pops (half-sine pressure pulses of 20–60 ms under 150 Hz before each word that follows a pause): VoiceBank+DEMAND test utterances, Spoken Wikipedia narrations, the music repair uses, 0.2.0 → 0.3.0; voiced frames thinned: the voiced 10 ms frames whose level under 250 Hz fell by more than 3 dB:

Pops peaking at 0.5, 1 and 2× the utterance's peak, the error to the clean speech taken away over each pop and the 150 ms after it (dB):

Use when: mic plosives (p, b, t) producing low-frequency thuds.
Not for: a kick drum: it is a pop to this (a jazz mix loses 3 dB under 250 Hz on 1.6 % of its frames, 0.2.0: 0.4 %); a live path that cannot wait 14 ms.

Part of @audio/denoise — the denoise family umbrella. This README is generated from the umbrella docs.

MIT © audiojs