Voice

Two separate things, each with its own switch and its own model. Dictation writes into the composer and stops there. Voice mode runs the whole exchange without your hands, and only on the phone interface.

Both are extra models, billed like any other, and most connections offer neither. Check Providers before expecting them.

Speaking instead of typing

Switch on Settings, Voice, Voice input and choose a model to transcribe with. Only models marked as transcription are offered; if the list is empty, mark one in Models and prices on its connection.

A microphone then appears in the composer, on a desktop as much as on a phone. It writes into the field rather than sending, so what was heard is read before it goes.

Voice mode

The phone interface gives speaking a screen of its own, and there it is a loop rather than a microphone button. You speak; it stops listening on its own once the room has been quiet for a moment; the turn runs; the answer is read back out loud; it listens again. Nothing is tapped between one question and the next, which is what lets it work with the phone face down on a table. One key stops whatever is happening, and pressing it while an answer is being read cuts it off and hands you the floor.

Reading aloud is its own switch, Settings, Voice, Read answers aloud, with its own model and a voice chosen by name. It is off until asked for: it is a second model again, billed like any other, and on most connections it is not available at all.

How long a silence ends your turn, and whether it listens again once it has answered, are both in Settings, Voice. Neither has a right value in the abstract: somebody who thinks mid-sentence needs three seconds where somebody dictating a list is cut off by anything over one, and a microphone that reopens by itself is a reasonable thing to switch off.

The orb is the control. There is no separate key: you press the only object on the screen, and its colour says whether a conversation is engaged. Muted is a screen waiting to be started, the accent is a microphone that will open again by itself once the answer finishes. Pressing while it reads takes the floor back.

It is drawn from the sound itself rather than animated on a timer, the microphone while you talk and the speaker while it answers, so a room it cannot hear leaves it still. That is the quickest way to tell a permission problem from a quiet one.

An exchange held by voice is an ordinary conversation. It is saved as it goes and it is in the list afterwards, every word of it readable, because a voice mode that kept no transcript would be the one part of the app you could not go back and check.

The sound never touches disk in either direction. A recording exists for the length of one request to your own transcription model; an answer exists as sound for as long as it is playing, and is never cached, since a cache of what you were told is a transcript kept by the one part of the app with no reason to hold one. On a shared instance an administrator can publish the transcription setup, which is usually the difference between a feature everybody has and one only they have.

It installs as a PWA. On iOS, Share → Add to Home Screen; on Android, the install prompt in Chrome’s menu. Installed, it runs without browser chrome and keeps its own theme colour.