Voice

Both directions run on the device.

Dictation

Speak instead of typing and the audio is transcribed locally by WhisperKit. The model (openai_whisper-base, about 145 MB) downloads the first time you use the feature.

The recording is deleted as soon as it has been transcribed. It is never stored, and never sent anywhere.

Reading replies aloud

Replies can be spoken using the operating system's own speech synthesiser, sentence by sentence as the reply generates rather than only when it finishes — so a long answer starts being read almost immediately.

Configure it in Settings → Speech: turn automatic reading on or off, set the speaking rate, and preview the voice.

No cloud involved

Neither direction uses a network service. Dictation does not go to a speech API, and reading aloud does not fetch a voice. Both work in aeroplane mode once the Whisper model has downloaded.