Voice and dictation

Dictation types what you say into the prompt bar, and the voice window listens while it is open and sends each sentence to the agent.

Bandicot listens only while you dictate or while the voice window is open. There is no wake word, and Bandicot never speaks back: replies are text on screen.

Click the mic or press ⇧⌘D, speak, and the words appear in the prompt bar for you to check and send.

When a speech model is set for Speech Recognition on the Rack, it transcribes on your Mac, up to 30 seconds at a time. Otherwise dictation uses macOS speech recognition. See Local models.

With a Rack speech model, press the mic again to stop early. macOS dictation uses your Mac's language. The mic shows only when Dictate is on and macOS speech or a Rack speech model is available. While the voice window is open the mic button takes you to the window instead of starting dictation, because only one thing can hold the microphone.

Press ⌥⌘T, say what you need, and each sentence goes to the agent as a message, marked with a small microphone.

The voice window needs macOS 14 or later. The first time, it asks to download the voice models, which then run on your Mac. A chat started by voice is named after what you said.

The mark's eyes show the state, and each sentence gets a dot: sent, done, queued for after this run, failed, or kept here and not sent. Close the window with its close button, ⌥⌘T or Esc; a click outside does not close it.

The window talks to the chat in front of you. Switch chats and your next sentence goes to the new one; the thought, the next-step line and the queue clear. ⌥⌘T works while Bandicot is in front.

Under your sentences the window can show the agent's latest thought, which you can click to show the reasoning in the transcript, and a next-step line with an arrow. The line ends with say yes or no when you can answer by voice, or answer this one on screen when you must use the card. Sentences waiting for later appear as small chips, Queued for later. When more than one voice is heard, a group of dots reads N voices; only yours is sent on hover. The ? button opens an explanation of every dot and colour.

If the voice models are missing, the window offers Download the voice models (529 MB), shows Downloading the voice models… with how much has arrived, and starts listening by itself when the download ends.

1. 1Press ⌥⌘T, or hold the mic button.
2. 2The first time, choose to download the voice models and wait for them.
3. 3Speak when the eyes show Listening.
4. 4Read your sentence in the window and the agent's reply in the chat.

Each sentence in the window gets a dot that says whether it was sent, queued, kept here or could not be sent.

Hover a sentence to see the words, or open the ? help to see them beside every sentence. A struck-through sentence is another voice and is never sent.

Say stop, cancel, wait, yes or no, in English or Spanish, to control the run without touching the keyboard.

A command must be a short sentence on its own, up to about forty-eight characters; please or por favor beside it is fine. Anything longer goes to the agent as an ordinary sentence.

stop and cancel do the same thing: they end the run and close any permission card or question that is open. They are the one command accepted before the window has matched the voice, and only while yours is the voice in the window.

wait holds the agent before its next step and the window shows Holding before the next step. Say yes or any new sentence to let it go on, or stop to end the run; a hold also lets go by itself after five minutes.

yes allows once and no denies, but only while a permission card is open; yes with no card simply lets a held run go on. Riskier asks must be answered on screen. A bare wait, yes or no with no run to answer is kept in the window and not sent, so it never starts a paid run. Any other sentence said while the agent works is sorted: a steer for now, a follow-up for later, or an aside kept in the window and not sent. To take back a change, use Versions; saying undo does not do it.

A sentence said while a run is still finishing is not lost. It waits, then starts the next run.

The voice window learns the first voice it hears and keeps a short sample of it on this Mac, so other voices are shown struck through and never sent.

The sample is about six seconds long, stays on this Mac only, and is never versioned or copied. Speak first when you open the window: Bandicot takes the first voice it hears as yours, and keeps it only after hearing it first again on a later start. Until then, another voice nearby may be taken for yours.

To start over, open the window's help (?) and choose Forget my voice. It answers Forgotten. It is learnt again next time., or No voice was stored. when there was nothing to forget. That's not me shows while the window is listening and can tell voices apart; it answers Forgotten. Speak now and it learns your voice. and learns the next voice it hears. If it fails it says Could not forget it. Try again.

If this Mac cannot tell voices apart, the window warns that anyone speaking near it is sent as you.

Mute the microphone in the voice window turns the microphone fully off, so macOS shows it as off and nothing is heard, while the window stays open.

Choose Unmute the microphone to listen again. Words said while muted are dropped.

In the voice window, speech becomes text on your Mac, and only that text goes to your model provider, like a typed message.

Audio from the voice window is never written to disk. With macOS dictation, the system speech service may be involved, so set a Rack speech model if you need dictation to stay on this Mac. See Privacy.

Most voice problems come from a missing macOS permission, an older macOS, or voice models that are not downloaded yet.

The window says what went wrong in one line. Fix that, then open the window again.