How to talk to a local AI on your Mac

Dictation gets dismissed as a convenience feature, which undersells it. Talking is how most people produce a first draft fastest — and the reason to do it locally is that your voice is unusually personal data to hand to a server.

Why local voice is different

Cloud dictation ships audio of you speaking to someone else's computer. Even where that is handled carefully, it is a recording of your voice, often capturing whatever else was audible in the room.

Doing it on-device removes the question. Nothing is uploaded, so there is no retention policy to read and no incident that can expose it.

Using it

1. Start dictating from the composer and speak normally — you don't need to slow down for it.
2. Send it like any other message; edit first if you want.
3. Have answers read aloud when you'd rather listen than read.

Where it fits and where it doesn't

• Good: long rambling context, first drafts, thinking out loud, hands busy.
• Good: listening to a long answer while doing something else.
• Poor: code. Reading punctuation and symbols aloud helps nobody — when an answer is all code the app tells you why it isn't reading it, rather than staying silent and looking broken.

Two things to know before you download for this

It is push-to-record, not a hands-free conversation. You hold to record, release, and the transcript lands in the composer where you can edit it before sending. There is no wake word, no always-listening mode, and no spoken back-and-forth where the app replies out loud and then waits for you to speak again. If what you want is a voice assistant you talk with, this is not that, and no amount of settings will make it that. What it is: a faster way to get a long, messy thought into the box than typing it, plus the option to have the answer read back.

Voice input needs a paid plan. The free tier — Nano and Lite — is text chat. Dictation and read-aloud are part of Pro, along with the other tiers and features. Worth saying plainly here rather than letting you find out after the download.

Getting a clean transcript

Dictation is transcription, not comprehension: it writes down sounds and has no idea what your project is called. Three things that actually help.

• Say the punctuation you care about. "period", "comma", "new paragraph" land as marks. Everything else arrives as one long run-on, which is fine for a prompt and bad for anything you plan to paste elsewhere.
• Spell unusual names once, then stop. Product names, library names and surnames are where transcription goes wrong. Fix it once in the composer before you send — the model reads what is in the box, not what you said.
• Record where it is quiet. On-device transcription has no cloud model to lean on for a second guess, which is the privacy trade working as intended. A noisy room costs you accuracy that a cloud service would have papered over.

Common questions