Audio processing

Capture runs at 48 kHz through a chain of Web Audio nodes. Every stage after the volume control is optional, and each one can be turned off on its own from Settings → Audio.

Mute sits last on purpose. Nothing upstream of it can defeat it, so muting is muting regardless of what the rest of the chain is doing.

Volume

A GainNode on the raw input. This is also the tap the noise gate measures against, which matters later.

RNNoise

Neural noise suppression, running as an AudioWorkletNode rather than on the main thread. It removes keyboard clatter, fans and background hum while leaving speech alone.

If the worklet fails to connect, the error is logged and the chain carries on without it rather than failing to produce audio at all.

An AnalyserNode at 2048 FFT with smoothing set to zero, feeding a GainNode. Zero smoothing means it reacts to the level it just measured rather than to a rolling average, which is what makes it track a voice that moves closer to and further from the microphone.

Compressor

A DynamicsCompressorNode, controlled by a single Amount slider from 0 to 100. The slider interpolates three of the five parameters:

Amount	Threshold	Knee	Ratio
0	-10 dB	40	1:1
50	-25 dB	22.5	10.5:1
100	-40 dB	5	20:1

Attack stays at 3 ms and release at 250 ms throughout.

Noise gate

Also an AudioWorkletNode, which is deliberate: a worklet keeps running when the window is hidden, and a gate that stops working when you tab away is worse than no gate.

The threshold is a percentage from 0 to 100, not a decibel figure. Audio below it is silenced.

The interesting part is what it measures. The gate takes its signal from the processing chain but its threshold reference from a tap taken before RNNoise, AGC and the compressor. Measuring a post-chain level against a threshold the user set while watching a pre-chain meter would quietly change what their number means.

Release time is configurable. Smoothing is 0.8 normally and 0.3 in eSports mode.

Push-to-talk turns the gate off

With push-to-talk selected the threshold is set to 0, which disables gating entirely. The gating is done by the mute node instead, since that is exactly what push-to-talk is. The noise gate slider is hidden in that mode rather than sitting there doing nothing.

Mute

A GainNode at zero. The state is synchronised with the server in both directions, so what other people see on your name matches what your microphone is actually doing, including when a moderator mutes you.

Output

A final AnalyserNode drives the level meters, and a MediaStreamDestination hands the result to WebRTC.

There is a separate monitor tap before the mute node, which is what loopback monitoring listens to. That placement means you hear yourself as others would, minus the mute.

eSports mode

One toggle in Settings → Voice that trades processing for latency:

RNNoise off
Noise gate bypassed
Push-to-talk enabled
Bitrate capped at 128 kbps
Opus packetisation set to 10 ms frames

The frame size is where most of the gain is. Shorter frames mean less time spent filling a packet before it can be sent.

Capturing system audio without also capturing Gryt's own output is not something the browser APIs will do. The desktop app runs a native binary that excludes Gryt's process tree, writes raw PCM to stdout, and hands it to the Electron main process. That forwards chunks over IPC to the renderer, where an AudioWorkletNode turns the stream back into a MediaStreamTrack WebRTC can send.

Devices

Input and output devices are chosen in Settings → Audio and remembered in localStorage along with volume, gate threshold and every toggle above. None of it is server-side, so it follows the installation rather than the account.

When something sounds wrong

Others hear themselves when you share audio. The capture is picking up your speakers or including Gryt's own output. Use the desktop app's native capture rather than a browser tab, or headphones.

Your voice cuts off at the start of words. The gate threshold is too high, or its release is too short. Watch the raw input meter next to the slider and set the threshold below where your quiet speech sits.

Your level pumps up and down. AGC and a heavy compressor setting fighting each other. Turn the compressor amount down first.

Nothing is transmitted at all. Check the mute state, then that the right input device is selected, then browser microphone permission. The level meter shows raw input, so a meter that moves while nobody hears you points downstream of capture, most often the gate.

For anything that looks like a connection problem rather than an audio one, the voice debugging page is the better place.