Dictation

Hold fn in any app and talk; let go, and clean text lands at your cursor, written by models that run on your Mac.

Dictate

1. Click where you want the text, in any app.
2. Hold fn (the 🌐 key) and start talking. A small pill at the bottom of the screen shows that it's listening, with a timer.
3. Let go. The words settle into clean text, and it's typed in at your cursor.

The first time, Saiki downloads its speech model once (about 220 MB); the pill shows "Getting dictation ready" and then "Dictation is ready: hold fn and talk".

• Dictation starts after you've held fn for a quarter of a second with nothing else pressed, so fn shortcuts like fn↑ or fn⌫ keep working.
• esc, or another key while you hold fn, cancels.
• A press shorter than a third of a second is ignored. One dictation can last up to 5 minutes.
• If nothing was heard, the pill says "No speech detected".

The pill never takes the keyboard or clicks from your app, and it follows you to other Spaces. With the Colorful Border look, a glow in the pill moves with your voice; with No Border, bars do. See Looks and motion.

Where the text goes

The text goes only into the app you started dictating in, at the cursor. Saiki types it in directly where the app allows, and otherwise pastes it and puts your clipboard back right after, without the text showing up in clipboard history.

If you've switched to another app by the time it's ready, nothing is typed anywhere. The text is copied instead, and the pill says "You switched apps, so it's copied: ⌘V to paste it".

Spacing and capitals fit what's already there: a space after the previous word, lower case in the middle of a sentence.

Terminals get one line, and never Enter

In a terminal (Terminal, iTerm, Warp, Ghostty, kitty, Alacritty and WezTerm), the text is always one line, because a new line there would run it. Dictation never presses ↵ or submits anything, in any app: the text is typed in, and running it is up to you.

Clean-up

After the speech model writes down what you said, a clean-up model tidies it. Pick how much in Settings → Dictation → Clean-up:

In code editors, terminals and search boxes, letters you spell out become one word ("C. D." becomes cd), and there's no full stop at the end.

The window's text as context

When you start, Saiki reads what's around your cursor through Accessibility: the app, the window's title, the field's placeholder, the text just before the cursor, and the window's visible text (up to about 2,000 characters). The clean-up model uses it to understand where you are, so a name or a folder on screen comes out spelled the way it's written there. No Screen Recording is needed.

That text is treated as reference only: the model never takes instructions from it. It's never read in password fields or in password managers, never leaves your Mac and is never logged.

Exact words: ⌥⌘Z

Not what you meant? Within 8 seconds of a dictation, press ⌥⌘Z and the cleaned text is swapped for exactly what you said. For a few seconds after the text goes in, the pill shows Use Raw as a reminder.

Models on your Mac

Everything runs on your Mac: your voice and your words are never sent anywhere. Choose the models in Settings → Dictation. The defaults are the smallest.

Speech Model

Clean-up Model

Each row shows where it stands: In Use, Installed, Built In, Download, a percentage while it downloads, or why it can't run on this Mac (such as "Needs macOS 26.2" or "Not in this build").

• ↵ on a model uses it, downloading it first if needed. ↵ again while it downloads stops the download.
• ⌘⌫ moves a downloaded model to the Trash. If it was in use, dictation goes back to the default.

Downloads come from fixed versions and are checked before use. A model loads when you dictate and is released a minute later, so it doesn't hold memory while you're not talking. If no clean-up model can run, Saiki still removes filler words and adds a capital and a full stop.

Your words

Instructions for every dictation

In Settings → Dictation → Your Words, Instructions for every dictation lets you write how you like your text, in your own words, for example: "use British spelling; never use exclamation marks." They're added to every clean-up, wherever you dictate. ⌘↵ saves, and Clear removes them.

The dictionary

Add Word… opens a small form:

• Word or phrase alone: a name or term to spell exactly. Dictation spells it like this and fixes near misses, one letter off.
• With Replace with: when you say the phrase, the replacement is written instead, such as "my email" becoming your address. Replacements apply even with Clean-up set to Raw.

Your words show in the list under it; choose one to edit or delete it.

Turn it off

Turn off the switch at the top of Settings → Dictation, or Dictation on the Features page. fn then does only what macOS makes it do.

Privacy

• Your voice, the transcript and the window's text stay on your Mac. Nothing is sent to any server.
• What you dictate and the window's text are never logged or saved.
• The window's text is never read in password fields or password managers.
• Dictation uses the Microphone and Accessibility permissions. See Install and permissions and Privacy and safety.