Nothing moves that shouldn't.
We make instruments for the human voice. Each one measures the signal, makes a single decision, and applies it once — to the one thing that was actually the problem. Everything else is left exactly where you put it.
What we build
Restraint is the product.
A voice arrives carrying things that are not the voice — a harsh ess, a room, a noise floor, a reverb tail belonging to a space you never chose. Taking one of those away without disturbing anything else is the entire discipline. It is also far harder than taking everything away, which is what most processors quietly do.
One decision, applied once
Measure, decide, act — in that order, exactly once. An instrument that keeps re-deciding mid-chain ends up arguing with itself, and the voice always loses that argument.
No fake analysis
Every meter shows something real. Every number we publish was measured on real material — and reported even where a kinder number was available.
Yours, permanently
An instrument you rent is an instrument that stops working the day you stop paying. Our licences are perpetual, and the free tier is licensed for commercial work.
The Clarity Series
Two instruments.
One takes away what the voice should never have brought with it. The other separates the voice from the room it was recorded in.

Sibilance
Built & validatedBest for: Transparent de-essing on lead vocals
A vocal de-esser. Transparent, level-independent de-essing that removes harsh sibilance without dulling the voice — plus a subtle analog finish.
Explore Sibilance
Thunder Clarity
In developmentBest for: Pulling a voice out of a noisy room
Real-time neural voice clarity. Separates the voice, its own reverb tail and the ambience into three faders you can rebalance live.
Explore Thunder ClarityThe engineering
Three lanes. No lane is trusted alone.
Human Engine Labs came out of aerospace, and it shows in how these instruments decide. The de-ess decision is made by three dissimilar lanes, each owning the call it is competent at and monitoring the others — the same command and monitor principle used in flight-control computers, so no single lane fault can cause a wrong action.
Drives de-essing: “is there sibilant energy here?” Fast, and it never hallucinates — but it cannot tell a real “s” from a broadband plosive, so it gets no vote on what the sound is.
It can tell them apart. It vetoes the command — holding back de-essing — only for a recognised plosive or fricative worth preserving.
Requires that veto to persist across consecutive windows, so a single confidently-wrong frame on a vowel can neither trip protection nor flash a false letter on the meter.
The AI never drives the cut. It only ever protects. And before any DSP touches the signal, non-finite input from a misbehaving host is flushed to zero and counted, so one bad sample can never poison the internal filters — a 200 ms NaN storm produces zero non-finite output, and de-essing self-heals the instant clean audio resumes.
The list
Be there for the first release.
Sibilance is built and validated. When it opens, the list hears first — and the people on it get the intro price. That is the only thing this form does.
Release notes and nothing else.
Apple Silicon and Intel. JUCE 8 · C++20.