Sibilance
v2.0Built & validatedThe ess is the only part of a vocal everyone agrees is a problem, and the only part every de-esser damages on its way past. Sibilance takes it out and leaves the rest of the voice precisely where the singer put it.
The idea
Everything else stays untouched.
A conventional de-esser hears a loud ess and pulls the whole top end down, taking the air, the consonants and the life with it. Sibilance measures where the ess energy actually sits and places a bank of adaptive minimum-phase notches on exactly that — so the two curves below are identical everywhere except the one band that needed help.
Detection
It knows an ess from a bright vowel.
This is where most de-essers fail. Loudness alone cannot tell a sibilant from a cymbal, a bright vowel, or a breath — so threshold-based designs either miss quiet esses or chew through everything above 6 kHz. Sibilance requires three things to agree at once.
HF-dominant
The energy has to actually be up in the sibilant band, not merely present somewhere in a loud passage.
Noise-like
Measured by spectral flatness. An ess is broadband hiss; a vowel is harmonic. This is what saves bright vowels and cymbals from the notch.
Unvoiced
No pitch. If the vocal cords are involved it is not an ess, no matter how bright the sound is.
Because none of those three cues is a loudness test, detection is level-independent: a quiet ess and a loud ess of the same character get the same treatment. On top of that, per-voice auto-calibration learns each singer's baseline brightness — which is why one dial works across a bright pop vocal and a dark baritone without being re-taught.
The controls
Five controls. No menus.
Hush
One dial does the job.
Turn it up to remove more. The number is the amount; the reading below it is live gain reduction in dB. Because detection is level-independent and calibrated per voice, this is usually the only control you touch.
Not sure where to start? Play a few seconds and tap Auto — the Tracking Assistant listens to how prevalent the sibilance actually is and sets a recommended starting point. Nudge from there.
The Monitor · Δ Listen
Hear exactly what you removed.
The graph is the sibilant band in real time. The scooped dip is what is being taken out. A vertical marker tracks the exact frequency the engine is working on right now — it follows the singer around.
Letters float up as sounds are detected: teal for sibilants being controlled, amber for consonants being protected. And Δ Listen solos only what is being removed — it goes silent on everything else, which is how you prove it is eating esses and not your vocal.
Silk · Crisp · Smooth · Warmth · Level
The finish, after the decision.
Silk puts an airy sheen back on top — applied after de-essing, so it can never re-harshen the esses you just tamed. Crisp restores the crisp top of the “s” that heavy de-essing strips: the antidote to a lisp.
Warmth is gain-matched analog saturation that colours the body and never the highs, in three voicings — Transformer for balanced gentle harmonics, Tube for richer even-order and a softly compressed feel, Tape for odd-order and a silky top. Smooth pulls down harsh resonances across ~1.8–16 kHz and quietly cleans steady broadband hiss. Level trims the output.
Routing & workflow
Where it needs to work.
De-ess the full stereo image, only the centre, or only the sides — Mid is how you keep a stereo double's sibilance centred. Drive detection from an external sidechain when you want another track making the call.
De-Pop catches the brief low-frequency thump of a plosive and high-passes only that instant, leaving sustained bass and the vocal fundamental alone. Breath ducks the quiet unvoiced moments between phrases and leaves the words untouched. Auto Gain is on by default, so the macros never change your level and bypass is a fair A/B. Two full A/B slots, and Copy to duplicate one into the other.
Measured
The numbers, reported honestly.
Validated end-to-end on realistic vocals — not on synthetic material, which would let us print better figures than these.
- Classifier accuracy
- On held-out real speech: s ~97% · sh ~94% · f-th ~76% · t ~70%. We publish these rather than the higher synthetic-only figures.
- Oversampling
- Adaptive 4× / 2× / 1×, chosen by host sample rate.
- Latency
- Reported to the host and verified sample-accurate; bypass is latency-compensated.
- Smoothing
- 20 ms parameter smoothing. Click-free stereo-mode and auto-gain switching.
- Integrity
- Input NaN/Inf guard ahead of all DSP. A 200 ms non-finite storm yields zero non-finite output.
- Formats
- VST3 · Audio Unit · CLAP · Standalone. macOS (Apple Silicon & Intel) and Windows.
Presets
Ten starting points.
Factory presets are starting points, not destinations. Your own presets save the entire state — every macro, the routing, Auto Gain.
Pricing
Bought once. Yours.
The free tier is a real de-esser licensed for real work — not a demo with the good parts switched off.
- The full single-authority engine
- Hush, Auto and Level
- Δ Listen and Auto Gain
- Licensed for commercial work
- Everything in Open
- Silk, Crisp, Smooth and Warmth
- Mid / Side and stereo modes
- De-Pop, Breath and de-lisp
- External sidechain, A/B, all ten presets
- Sibilance Pro
- Thunder Clarity at 1.0
- Every future point release