Keyboard noise on voice chat: what RNNoise actually removes (we measured it)
Why key clicks slip through browser noise suppression, how RNNoise handles them, and our measurement — speech level within 0.7 dB, keyboard clicks down 9.3 dB.
On this page
Short answer: Keyboard clicks are short, sharp bursts, and classic noise suppressors are built for steady noise like fans and hum — so clicks often get through. RNNoise, a small neural network running on your own device, tracks speech instead of noise. In our measurement it kept speech within 0.7 dB of its original level while turning keyboard clicks down by 9.3 dB.
If your teammates can hear every key you press, the problem is rarely your microphone alone. It is usually the noise suppressor in the middle, and the kind of noise it was designed for.
Why keyboard clicks slip through
Most real-time noise suppression works by estimating the background noise and subtracting it. That works well for noise that is stationary: a PC fan, an air conditioner, electrical hum. The estimate settles in, and the suppressor removes roughly the same thing every frame.
A key press is the opposite. It is a transient — a few milliseconds of broadband energy, then silence. By the time a slow-moving noise estimate notices it, the click is already over and has gone out to everyone in the room. In our own testing, the browser's built-in suppressor (the noiseSuppression constraint in Chromium's WebRTC audio pipeline) sometimes let clicks through and, in the other direction, sometimes clipped the start of words.
What RNNoise does differently
RNNoise is an open-source noise suppressor from Xiph.Org, the people behind the Opus codec. Instead of modelling the noise, it combines traditional signal processing with a small recurrent neural network that has learned what speech looks like:
- Audio is processed at 48 kHz in 10 ms frames.
- For each frame, the network predicts a gain for 22 frequency bands — how much of each band is voice and how much is not.
- Because the decision is made frame by frame from what speech should look like, a click that does not resemble a voice is turned down even though it never "settled in" as background noise.
It is small enough to run in real time inside a browser tab as a WebAssembly AudioWorklet, so your microphone is cleaned on your device, before anything is sent.
How we measured it
We wanted a number, not an impression. The test used the exact build that ships in Big Game Area — the same compiled worklet and WebAssembly module — rendered offline in the browser (OfflineAudioContext) so the run is repeatable:
- Speech: synthetic syllables with a moving pitch contour, standing in for a talking voice.
- Keyboard: a sharp click every 0.65 seconds, layered on top.
- Room: low-level background noise underneath both.
We then measured the level of each part separately before and after RNNoise:
| Band | Raw | After RNNoise | Change |
|---|---|---|---|
| Speech (syllable windows) | −26.9 dB | −27.7 dB | −0.7 dB |
| Keyboard clicks | −23.5 dB | −32.8 dB | −9.3 dB |
How to read this: a change under 1 dB in speech level is hard to hear, so your voice comes through essentially untouched. A 9 dB drop is large: as a rule of thumb, about 10 dB reads as "half as loud", so the clicks go from competing with your voice to sitting well under it.
Limits of this test
This is a synthetic signal, not a recording of a real person on a real keyboard. Mechanical switches, desk resonance and microphone placement all change the result. Treat 9.3 dB as what the algorithm does to a click-like transient under controlled conditions, not as a guarantee for every setup. We plan to repeat the measurement with real recordings and will update this post when we do.
Five fixes that help with any voice app
Noise suppression is the last line of defence. These help regardless of what software you use:
- Move the microphone, not the keyboard. A headset boom mic a couple of centimetres from the corner of your mouth picks up far less keyboard than a desk mic aimed at your whole setup.
- Point a desk mic away from the keys. Most desk mics are directional (cardioid): they hear best in front and least from behind. Put the keyboard behind or to the side of the microphone.
- Lower the input gain and speak closer. If your input level is set very high, every small sound is amplified with your voice. Less gain plus a closer mic improves the voice-to-click ratio before any software runs.
- Damp the keyboard. A desk mat absorbs low-frequency thumps; O-rings or quieter switches reduce the click itself.
- Use push-to-talk in clutch moments. Nothing gets transmitted between callouts.
Turning it on in Big Game Area
RNNoise is on by default. In a room, open Audio settings to switch between three modes:
- Strong (RNNoise) (default): echo cancellation, automatic gain and RNNoise.
- Standard (browser): the browser's own noise suppression, kept for comparison.
- Off: echo cancellation and automatic gain only — your raw microphone.
If your browser cannot run the AudioWorklet (for example when WebAssembly is blocked), the call does not drop: the raw microphone is passed through and the settings panel tells you so.
Nothing needs to be installed and there is no paid add-on behind it — the model runs in the page, served from our own servers.