De-esser: tame harsh "s" sounds.

At a glance

Why esses get harsh

Sibilants are noise: the hiss of air through teeth, with most of their energy between about 4 and 10 kHz. A microphone close to the mouth, a bright condenser microphone, a presence boost in an EQ, heavy compression, or a podcast processed to sound loud on phone speakers all push that region up, and the esses start to cut through, spit and distort, worst of all on earbuds. Turning the treble down fixes the esses and dulls everything else. A de-esser turns the top end down only while an ess is happening.

How this de-esser works

First it finds where this voice's sibilance sits. It looks at the moments that are mostly top end and finds the peak of their spectrum, usually around 5 to 7 kHz for a deep voice and higher for a light one. The crossover is placed under that peak, between 3.5 and 8 kHz, so the body of the voice below is never touched. The sound is split into the part below the crossover and the part above, with a zero-phase filter run forwards and backwards, so the two parts add back to the original exactly.

Then it decides, every 5 milliseconds, whether an ess is happening: if more than 45% of the sound's energy is above the crossover, it is a sibilant. Vowels sit well under 10%, so a bright voice is not mistaken for one. The "Which sounds" setting moves that line. During an ess the top band is turned down by the amount you choose, easing in over 15 points above the line so borderline sounds get less. The gain changes along a 6 ms curve centred on each moment, so it starts turning down just before the ess arrives and lets go just after, without the pumping of a slow compressor. Everywhere else the output is not just close to the input but identical, sample for sample.

The check behind this page runs two voices reading "She sells sea shells by the sea shore" and more. It requires the esses' top end to come down by the setting within 1 dB (measured 5.6 to 5.8 dB at 6), within 1.5 dB at 12 (10.7 to 11.3), and the vowels to stay within 0.2 dB, which they do to three decimals. Every sample away from an ess must be the input's, and "only what is removed" plus the de-essed file must add back to the original.

How much to use

Start at 6 dB and listen on headphones or earbuds, where sibilance is worst. If an ess still spits, go to 9; if the speaker starts to sound lispy, with "s" turning to "th", go back down. Tick "only what is removed" to hear the cut on its own: it should be nothing but hiss. If you hear the body of words in it, use a lower setting or "Only the harshest esses". A de-esser goes after the EQ and before the final loudness, which is the order enhance speech uses. In a DAW the same job is done by FabFilter Pro-DS, Waves Renaissance DeEsser or Logic's DeEsser 2; this page does it to a finished file, in the browser, with the same length out so a video stays in sync.