Stop recording when you stop talking

Agents: claude, codex, opencode · On: phone, desktop · optional

A recording ends by itself after a set number of seconds of quiet, so you do not have to reach back for the stop button. The window is a number in the cog, and 0 turns it off.

Why you would use it

You are dictating one-handed and the phone is at your mouth, not in front of your eyes. Three seconds of quiet ends the recording and starts the transcription. Set it longer when you think mid-sentence, or to 0 when you want the recording to run until you tap stop.

How to use it

Tap the cog (#settings-open) in the top right.
Find the Voice group (#settings-voice).
Put a whole number of seconds in Stop after silence (#voice-silence), the field with sec beside it. 0 to 30; the row's help says 0 turns it off.
The value saves as you leave the field. The next recording uses it.

What you see

The right end of the Voice title (#voice-hint) shows Saving…, then a green check and Saved. A bad value shows a whole number of seconds, 0 to 30 there in red, and the toast the silence window is a whole number of seconds, 0 (off) to 30, and nothing is saved.

While recording, the ring around the stop button fills toward the 120 second cap. There is no separate countdown for the silence window.

Options and settings

Option	Default	What it changes
`#voice-silence` (`voice.silenceSeconds`)	`3`	Seconds of quiet that end the recording. `0` is off
Range	`0`–`30`	Anything else is refused by `lib/settings.js` with `voice.silenceSeconds must be a whole number from 0 (off) to 30`
Hard cap	`120000` ms	`MAX_CLIP_MS` in `public/voice.js`. Not configurable

Limits and known gaps

The setting lives in settings.json on the box, not on the phone, so it is the same for every device signed into this box.
The silence stop is decided by a level threshold (SPEECH_RMS, 0.02) that has never been measured against a real phone microphone in a real room. It is only allowed to decide when a recording stops, never whether anything was said, so a soft voice under the threshold ends the recording early but the audio is still transcribed.
A window of quiet with no speech in it at all also ends the recording. If that happens in under one second, the clip is dropped without being sent; otherwise it is sent and comes back as heard nothing.
If the browser's audio graph cannot be read (a suspended AudioContext), the level counts as speech and the recording runs to the 120 second cap instead of being cut short. A graph that wakes mid-recording gets its silence stop back from that moment.
The cap applies whatever the setting says, including 0.