Noise Gates

A noise gate makes one decision, hundreds of times a second: is this signal worth hearing right now? Get the decision logic right and the gaps between your words go silent while your voice passes untouched. Get it wrong and it chews the ends off your sentences. Here's the difference — and why the best "gate" for speech isn't really a gate at all.

What a noise gate is

Gates belong to the same family as compressors, but they work the opposite side of the threshold. A compressor acts on signal above a threshold, turning loud moments down. An expander acts on signal below it, turning quiet moments further down. A downward expander with a 1:2 ratio pushes a signal that falls 5 dB under the threshold down by 10 dB; make the ratio steeper and steeper and you approach the limit case: anything under the threshold is muted entirely. That limit case is the classic noise gate.

The appeal is obvious. Most of the noise in a spoken recording — keyboard clatter, mouse clicks, chair creaks, the room itself — is quieter than the voice. Draw a line between the two levels, silence everything under the line, and the recording sounds like it was made in a much quieter room. The catch, as we'll see, is that "silence everything under the line" is a brutal operation to perform on something as fluid as speech.

A short history: the Kepex and the biggest drum sound in pop

Gating became a studio staple in the early 1970s with Allison Research's Kepex — the "Keyable Program Expander," introduced in 1971. Two things about it mattered. First, it was an expander with adjustable behavior rather than a bare on/off switch. Second, it was keyable: the gate on one channel could be triggered by a different signal entirely, which let engineers do things like tighten a bass guitar by keying it from the kick drum. Racks of Kepexes became standard studio furniture, and in live sound the same idea solved a chronic problem: tom-tom mics on a drum kit spill everything — cymbals, snare, stage wash — so engineers gated them to open only when the tom itself was actually hit.

The technique's most famous moment was an accident. At London's Townhouse Studios around 1980, engineer Hugh Padgham was working on Peter Gabriel's third solo album when a heavily compressed talkback mic in the drum room, chopped off abruptly by a gate, produced an enormous, unnatural drum ambience — a room sound that swelled and then vanished instead of decaying. The effect, refined and pushed harder, became gated reverb, and when Phil Collins used it on "In the Air Tonight" in 1981 it defined the sound of a decade of pop drums. It remains the best-known example of a noise-suppression tool becoming an instrument in its own right.

Why gates exist

The core justification is that noise is most audible when nothing else is happening. While you're speaking, your voice masks most of the junk underneath it. The moment you pause, the floor is exposed: the hiss, the fan, the flatmate's footsteps, your own fingers on the desk. A gate concentrates its effort exactly where the problem is — the gaps.

There's a second, sneakier reason. Any processing that raises average level also raises the floor between words. Compression with makeup gain can lift room tone by several decibels, turning noise you'd never noticed into noise everyone on the call notices. A gate (or expander) placed in the chain keeps the between-phrase floor down so the rest of the chain has clean silence to amplify. And unlike spectral noise reduction, which needs noise to be statistically steady, a gate doesn't care what the noise is — a cough from the next room and a single desk thump are treated exactly the same, as long as they're quieter than you.

How a gate actually works

A gate can't act on the raw waveform, which swings positive and negative thousands of times a second. It acts on an envelope follower: a rectified, smoothed running measure of how loud the signal is right now, typically averaged over a few tens of milliseconds. The envelope is compared against the threshold, and a small state machine decides what the gain should do. Its behavior is set by a handful of time constants:

• Attack — how fast the gate opens once the envelope crosses the threshold. This must be very fast for speech, because the information-carrying part of a word onset is over in milliseconds.
• Hold — a minimum time the gate stays open after the envelope falls back below the threshold, so brief dips inside a phrase don't trigger a close.
• Release — how fast the gain ramps back down once the hold expires. A slow release turns the close into a fade instead of a cut.
• Hysteresis — many gates open at one threshold but only close at a lower one. A signal hovering near a single threshold would otherwise flutter the gate open and shut many times a second, a fault engineers call chattering.

The gate's whole character lives in these numbers. The threshold decides what counts as signal; the time constants decide how gracefully it changes its mind.

Why hard gates sound terrible on speech

Speech does not stop; it decays. The tail of a spoken word slides down through 20 or 30 dB over a few hundred milliseconds, and somewhere on that slide it crosses any threshold you set. A hard gate — one that mutes — amputates the tail at that crossing. Sentences end with a tiny, unnatural snap instead of a natural fade, and once you hear it you can't stop hearing it.

The second failure mode is the floor itself. When a hard gate closes, the room tone doesn't get quieter — it gets switched off, then switched back on with the next word. Listeners are remarkably tolerant of steady background noise and remarkably intolerant of background noise that blinks. Breaths make it worse: a breath before a phrase is often loud enough to open the gate on its own, so the listener hears silence, then an amplified inhale, then the sentence. The fix used in modern speech processing is to abandon muting altogether: expand downward to a quiet-but-alive level instead of zero, open fast, hold through the natural gaps in a phrase, and release gently. Done that way, gating stops being an effect and becomes invisible.

In DeskBroadcast

DeskBroadcast's gate is built on exactly that principle: it is a downward expander, not a mute. When the level falls below the threshold, gain is reduced to about −30 dB rather than to silence, so the room stays faintly, naturally present instead of blinking on and off. The detector uses a 50 ms envelope, and the time constants are tuned for speech: a 2 ms attack that opens fast enough to never clip a word onset, a 150 ms hold after the level drops that bridges the gaps between words, and a 120 ms release that fades the floor down rather than cutting it. The one decision left to you is the threshold, adjustable from −70 to −30 dBFS to match your room and mic. Its job in the chain is specific: killing desk taps, mouse clicks and keyboard noise between phrases — the impulsive junk that lands while you're not talking. (Clicks that land while you are talking are a different problem, handled by the click suppressor, and steady noise under your voice belongs to spectral noise reduction and AI noise removal.)