Sibilance, and why your mic exaggerates it
Say the word "seaside" out loud. Those two s sounds are sibilance: bursts of high-frequency hiss your mouth makes on s, sh, ch, z, and hard t sounds. Unlike the tone of your voice, which lives mostly below 4 kHz, sibilance concentrates its energy roughly between 2 and 12 kHz, right where human hearing is most sensitive.
Microphones make it worse. Close miking, the presence boost built into most vocal mics, and bright condenser capsules all push that band up, so a voice that sounds fine in the room comes through the mic with every s sharpened into a little dart. Boosting the highs with an EQ for clarity sharpens the darts further. That is the problem a de-esser exists to solve.
A compressor that listens to one band
A de-esser is not a new kind of effect. It is a compressor whose detector is deaf to everything except the sibilant band. When energy in that band spikes, it turns the signal down; the rest of the time it does nothing at all. Four Beginner controls define it, and they are the actual parameters on the Patchd de-esser node.
- Threshold (default -28 dB). The loudness line, but drawn only for the sibilant band. Normal speech energy in that band sits below it; an s sound spikes above it and triggers reduction.
- Range (default 8 dB). The maximum reduction the de-esser is allowed to apply, from 0 up to 20 dB. This is the safety rail that keeps an aggressive setting from lisping your voice.
- Frequency (default 6500 Hz). The center of the band the detector listens to, adjustable from 2000 to 12000 Hz. Most voices hiss loudest somewhere near the default.
- Bandwidth (default 1.5x). How wide the listening band is around that center, from a narrow 0.5x up to a very wide 5x.
The node draws all four on its own display: the band and its center across the top, Threshold as the THR line and Range as the RANGE line on the meters below. Here it is in the middle of an s.
For most voices the setup is: leave everything at default, speak a few s-heavy sentences, and lower the threshold until the harshness softens. To aim the band by ear, the node's LISTEN button plays only what the detector hears, and DELTA plays only what is being removed, which should be little more than the hiss itself. If your voice still sounds exactly like you, only less spiky, it is set correctly. A de-esser you can hear working is a de-esser set too hard.
Split-band or broadband
The Mode toggle decides what gets turned down when sibilance trips the detector, and it is the biggest character difference between de-essers.
- Split-band (the default). Only the top of the signal is turned down. The node splits your voice at the detection band's low edge, near 5.3 kHz at the defaults, and applies the reduction to everything above that split. Below it the reduction tapers off at 12 dB per octave, so even a full 8 dB clamp costs only about 1 dB at 3 kHz, and the low and mid content of your voice passes through essentially untouched, even in the middle of an s. That is why split-band is the transparent choice.
- Broadband. The whole signal ducks whenever sibilance hits. It is the simpler, older design and it can sound perfectly natural in small doses, but every s briefly pulls down your low end too, which reads as a subtle dip in body on s-heavy phrases.
In either mode, Ratio (default 6:1) sets how hard the compression works once the band crosses the threshold, adjustable from a gentle 2:1 up to 20:1. The 6:1 default is deliberately firm: sibilant bursts are short and spiky, so a de-esser leans harder than the 3:1 you would use on a whole voice.
Why it goes after the compressor
Chain order matters here more than for most effects. A compressor raises your average level: quiet syllables come up, makeup gain lifts everything, and sibilance, which was already the sharpest thing in the signal, gets pushed up right along with it. A voice that barely needed de-essing before the compressor often clearly needs it after. So the de-esser sits downstream of the compressor, where it can clean up exactly what compression emphasized. The general rule, covered in chain order, still applies: cleanup first, character in the middle, and the de-esser right after the dynamics that feed it.
The timing and the math
The advanced controls are the standard compressor timing pair, tuned for sibilance: Attack (default 1 ms, 0.1 to 20 ms) and Release (default 60 ms, 10 to 500 ms). An s burst lasts maybe 50 to 150 ms, so the de-esser has to clamp almost instantly and let go before the next syllable. At 48 kHz, a 1 ms attack means the reduction is about two thirds of the way in 48 samples after the detector fires, and a full 8 dB clamp is within a tenth of a decibel of its depth about 5 ms in.
The detector itself watches one slice of the spectrum. The bandwidth multiplier is the ratio of that slice's top edge to its bottom edge, split evenly either side of the center on a musical, logarithmic scale:
f_low = fc / √bw, f_high = fc * √bwAt the defaults, the detector listens from about 6500 / 1.22 = 5307 Hz up to 6500 * 1.22 = 7961 Hz, where 1.22 is the square root of 1.5. Your s sounds have a zip code, and now the detector has the map.
When the level inside that band crosses the threshold, the gain reduction follows the usual compressor law, with one extra clamp: the Range control caps it.
reduction_dB = min(Range, (1 - 1/R) * (bandDb - T))At the defaults (threshold -28, ratio 6:1, range 8), a sibilant burst peaking at -16 dB in the band is 12 dB over, which asks for (1 - 1/6) * 12 = 10 dB of reduction, and Range clamps it to 8. Ten decibels of ambition, eight of permission.
And the reduction does not snap off the instant the s ends. It decays smoothly on the release time, the same exponential smoothing every compressor uses:
g = target + (g - target) * e^(-1 / (0.060 * fs))With a 60 ms release, the gain drifts back toward unity, shedding about two thirds of the remaining reduction every 60 ms. Long enough to be graceful, short enough that the next syllable never meets it.
One cross-reference worth knowing: the Patchd compressor has a sidechain low-pass on its detector, and it is the mirror image of this whole node. Pull a compressor's sidechain low-pass down and the compressor stops reacting to sibilance, so s sounds no longer pump the level of your entire voice. That is a compressor taught to ignore sibilance. A de-esser is the opposite lesson: a compressor taught to hear only sibilance. Same machinery, opposite filter, and a well-tuned voice chain often uses both.
De-essing in Patchd
In Patchd the de-esser is a free node available on every channel. Open the effects rack, add it from the + Add Node menu, and drop it in right after your compressor; dragging reorders it if the position needs to change.
The defaults described here, threshold -28, range 8 dB, frequency 6500 Hz, bandwidth 1.5x, split-band at 6:1, are the node's actual starting values, so most voices only ever touch the threshold. You can try it yourself: Patchd is in open beta now, and if you clear the terms your invitation arrives straight away.