@audio/dynamics-deesser

import deesser from '@audio/dynamics-deesser'

Sibilance reduction. An 's' is told by its shape, not its level: the band where sibilance lies (over split) is measured against the voice body (its formants, under 3.5 kHz), in dB, as the dbx 902 compares its high band with the full band. The same 's' is caught on a quiet recording and a loud one, and a cymbal in a mix, which sits over the mix's body, is left alone. How far the band rises over threshold sets the cut, held within range, the Range of the dbx 902, FabFilter Pro-DS and Waves Sibilance. A pause is measured against the voice's running level, so hiss between words is not taken for an 's'. The sound is read lookahead ms ahead of what is cut (a declared latency, so a host aligns it): an 's' sets its level within a few ms, and a cut that waits for it leaves its onset whole. mode applies the cut: split (default) to the band over split alone, x − (1 − g)·HP(x) with HP linear-phase (±1.5 ms), so x − HP is exactly the rest: no band is lifted at the split, and with no 's' the input passes sample for sample; band to a linear-phase band at fc, Q wide; broadband to the whole sound.

import deesser from '@audio/dynamics-deesser'

deesser(data)                                    // at any recording level; aligned (a stream is `lookahead` late)
deesser(data, { mode: 'band', fc: 7500, Q: 1 })  // only that band moves
deesser(data, { threshold: 3, range: -6 })       // fewer esses, shallower

Against iZotope RX 12 De-ess, node bench/rx/deess.mjs in audio (2026-10): VoiceBank test speech (2 speakers) and ten Spoken Wikipedia narrations, 21.4 min; its sibilants (active frames with half their energy over 3.5 kHz, found on the clean speech), 1702 of them, made harsh by their 4–10 kHz band alone, 4–12 dB up; the clean speech the target. Every setting chosen on 28 other speakers and ten other narrations, by SNR to the clean speech; RX tuned: Classic, threshold −9, Fast, cutoff 4 kHz (defaults: −12, 2.5 kHz). Over the sibilants, the error to the clean speech taken away and their 4–10 kHz band left over the clean's (median); over the reels, SNR to the clean speech; and the clean speech through it: SNR to itself, the share of its other active frames (vowels, voiced consonants) moved by over 1 dB, its own sibilants' 4–10 kHz cut (median):

0.3.0 cut late (with no look-ahead an 's' sets the level its own cut answers, so its first ms pass), on the whole sound or on a bell at fc, and no deeper than 6 dB, under the 4–12 dB a harsh 's' carries. node scripts/deesser.js runs it over speech, singing and music and reads, per 10 ms frame, how far the 4–10 kHz band and the 0.3–4 kHz voice body moved: on 's' frames (active, half their energy above 3.5 kHz) the cut, elsewhere the share of frames moved by more than 1 dB. VoiceBank test set (824 utterances), 10 narrations (60 s each), VocalSet:

Music ("Vibe Ace", Brahms, the Nutcracker, a trumpet): no frame moved by 1 dB in either mode. A ride cymbal 18, 12 and 6 dB under "Vibe Ace", the mix at −16 LUFS, its 4–10 kHz band cut: split, a median 0 dB (90th percentile 0, 1.1 and 4.4 dB); band, 0 dB (0, 0.8, 3.0). { mode: 'band', fc: 7500, Q: 1 }: 1.7 · 5.7 · 8.0 dB on the VoiceBank test set, 1.2 %.

Use when: harsh 's' / 'sh' in a voice, bright vocal takes; mode: 'band' for a lighter hand on one band.
Not for: broadband brightness: use an EQ. Generic compression: use compressor. A path that cannot wait 5 ms: lookahead: 0 (1.5 ms, the split's own).

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

MIT © audiojs