Click & Transient Suppression

Steady noise is easy: learn it, subtract it. A keyboard click is the opposite problem — over before any learner can learn it, and spectrally close enough to a consonant that removing one risks removing the other. This is the story of impulsive noise, why fixing it mid-speech is genuinely hard, and the small delay that makes it possible.

The opposite of steady noise

Every noise tool makes a bet about what noise is. Spectral noise reduction bets that noise is stationary — statistically the same from second to second — and wins against fans and hum by averaging the floor over time and subtracting it. An impulse breaks that bet completely. A keyboard click lasts a few milliseconds; by the time any averaging process has noticed it, it's gone, and it never repeats identically. To a stationary-noise tracker, a click is indistinguishable from signal.

A gate fails for the mirror-image reason. Gates pass everything above a threshold, and a click near your mic is usually louder than your voice, not quieter. Between phrases a gate's floor attenuates clicks nicely — but the click that lands while you're speaking sails straight through an open gate. Impulsive noise therefore needs its own detector: something that reacts to suddenness rather than to level or to statistics.

A short history: rescuing records

Declicking was born in audio restoration. The great archives of 20th-century sound — 78s, acetates, early vinyl — are riddled with impulses: dust in the groove, scratches, pressing faults. For decades the fixes were physical and manual: washing discs, and in tape-transfer workflows literally cutting out damaged instants. The 1980s made the problem computational. Once a recording is samples in memory, a click is just a short run of corrupted values, and it can be repaired by interpolation — modeling the signal on either side of the damage and resynthesizing the missing few milliseconds so the waveform passes seamlessly through where the click was. Systems of that era, most famously CEDAR (born from restoration research at Cambridge in the late 1980s) and Sonic Solutions' NoNoise, made this a professional discipline and remain the reference point for the field.

Two ideas from that era carried forward into every modern suppressor. First, detection and repair are separate problems: find the impulse with one algorithm, fix it with another. Second, the repair should be local — the click damaged a few milliseconds, so only those milliseconds should change, and they should be rebuilt from the healthy signal around them rather than simply silenced. Restoration tools, though, have a luxury a live voice chain doesn't: the whole recording, offline, with unlimited time to detect, model and interpolate. A microphone chain for a live call has to do its version of the same job in tens of milliseconds, on audio it hasn't heard yet.

Why a click mid-word is hard

Here's the genuinely difficult part, and it isn't detection speed — it's ambiguity. Speech is full of legitimate transients. A plosive consonant ("p", "t", "k") is a burst of broadband energy. A voiced word onset is a sudden low-frequency swell. For roughly the first 10 ms, a keyboard click and a "t", or a trackpad thump and the start of a vowel, look nearly identical to any detector: energy jumps, spectrum spreads, and the frame simply doesn't contain enough evidence yet to say which it is.

An over-eager suppressor that fires on every jump will duck your consonants — turning crisp speech mushy, which is worse than the clicks. An under-eager one is useless. The information that separates the two cases exists, but it lives a little later in time: a click decays — the mechanical event is over in milliseconds and its energy collapses — while a speech onset sustains, blooming into a vowel that carries on for a hundred milliseconds or more. The detector's problem is that it must decide at the jump, but the answer only arrives after it.

The look-ahead principle

The solution is disarmingly simple: don't decide at the jump. Delay the output slightly, so that when the algorithm is deciding what to do with a suspicious moment, it can already see what happened next. This is look-ahead — the audio equivalent of reading to the end of the sentence before judging a word. If the burst has collapsed a couple of frames later, it was a click: duck it, retroactively, in the still-delayed output. If the energy sustains, it was you: touch nothing.

Look-ahead is an old trick in audio — broadcast limiters have long delayed their output a few milliseconds so gain reduction can begin before a peak arrives rather than after it. Applied to click detection, it converts an impossible classification problem into an easy one: instead of asking "what is this jump?", the algorithm asks "did this jump survive?", and that question has a reliable answer.

The price is latency. Every millisecond of look-ahead is a millisecond between your lips and your listeners. For live conversation the budget is tight — tens of milliseconds pass unnoticed, but not hundreds — so the craft is in extracting a confident decision from the smallest possible peek at the future.

Repairing instead of muting

Detection is half the job; the intervention is the other half. Muting the whole signal for the duration of a click punches an audible hole in the voice — a dropout, often more distracting than the click. The restoration lesson applies here: the click only corrupted part of the signal, so only that part should be repaired. In a frequency-domain chain, that means identifying which frequency bins the impulse actually lifted and pulling those bins back toward where the voice's spectrum was just before the event, leaving the untouched bins alone. The voice continues through the event; the click is skimmed off its surface. Done well, the artifact isn't silence — it's just a click that seems to have happened much further away.

In DeskBroadcast

DeskBroadcast's click suppressor runs inside the same STFT as noise reduction, watching every frame for suddenness: a transient is flagged when high-frequency power (above 2 kHz) jumps to 3× its smoothed envelope. What happens next depends on the jump's spectral tilt. Jumps with a high HF/LF tilt are ruled clicks outright — keyboards and mouse buttons live there, and speech doesn't. The hard case is a low-frequency-dominant jump: a trackpad's chassis-conducted thump measures the same spectral tilt (about 0.01) as a voiced speech onset, so tilt alone can't decide. Those are resolved with two blocks of look-ahead — thumps collapse within about 20 ms, speech sustains — and that look-ahead is most of the stage's ~43 ms latency. When it does duck, it pulls only the frequency bins that jumped back down to the pre-click spectral envelope, so a click landing mid-word dulls to a distant thud instead of punching a hole in your voice. And it ducks once per event: the detector re-arms only after the spectrum settles, so one keystroke can't trigger a flurry of gain changes.