Voice dictation for macOS 26 and later · Apple Silicon · Menu bar · Nothing leaves the Mac
Hold a key, say it badly, and the sentence you meant lands at your cursor.
Hold the button and read this out loud, fillers and all: “um so i think we should uh meet on tuesday, actually make that wednesday”
This strip is a stand-in. It listens through your browser's own speech recognizer, which in most browsers sends your audio to that vendor's servers. Typing Optional does not: recognition runs on the Neural Engine and cleanup runs on an on-device model, on your Mac, with no request made at any stage. The cleanup pass you see here is a small copy of the real rules, written to run in a page.
What you said. What you meant.
Cleanup reads the whole utterance before it writes anything. That is what lets it drop the false start at the beginning and keep the correction you made four words later.
Heard
um so i think we should uh meet on tuesday actually make that wednesday
Inserted
I think we should meet on Wednesday.
- 3 fillers removed
- Self-correction resolved: tuesday becomes wednesday
- Capitalized and punctuated
The wait is shown, not hidden
Something appears within 150 milliseconds of the key coming up, on every dictation, before there is any result to report. Then a bar, with an estimate fitted to the throughput measured on your own machine and biased pessimistic, so the work usually finishes at or before the number you were shown.
Skip hands you the raw transcript immediately. And when cleanup comes back with substantially less text than it was given, the app throws that result away and inserts the raw transcript instead: a cleanup pass that summarizes has stopped tidying, and losing your words quietly is the worst thing it could do.
When insertion is refused
Some fields will not take synthetic text, and a password field never will. The transcript goes to the clipboard and the app says so. Nothing you said is discarded.
Correct it once. It stays corrected.
Names, jargon and product terms are where dictation visibly fails, and nobody fills in a dictionary by hand.
So the app watches what happens to the text after it lands. When you replace a word it inserted with one that sounds like it, that is evidence it misheard, and your spelling becomes a candidate term. Recognition is biased toward it from then on.
Every term is listed with where it came from, and any of them can be edited or thrown out. Reject one and it is not proposed again.
When it does learn something a panel says so, on screen, at the moment it happens: the term, where it came from, and a countdown that runs in the open before it dismisses itself. Hovering it stops the clock, because reading a list should not be a race.
Dictated
Hey Derek, I pushed the fix to the branch.
The app inserted “Derek”. Put the right spelling in and correct it, the way you would have anyway.
The dictionary
- Nothing learned yet.
It learns only from text you changed. Text you left alone teaches it nothing, which is the rule that stops it reinforcing its own mistakes.
The whole thing runs on the machine in front of you.
Four stages, all of them on your own hardware. The network lane under them stays empty, because nothing is put into it.
Capture
The key goes down and the microphone opens. The app does not have to be in front, and you do not change windows.
Recognize
Parakeet TDT v3 on the Neural Engine, through FluidAudio, which exposes the vocabulary boosting the learned dictionary needs.
Clean up
Apple Foundation Models, on device. No model to download, nothing to sign in to, streaming as it goes.
Insert
The text lands in whichever app was in front when you started talking, or on the clipboard if that app refuses it.
Nothing to opt out of
There is no telemetry, no account, no sync and no crash reporter. Turn off Wi-Fi and the product is unchanged, which is the only demonstration of a privacy claim worth anything.
Memory you can actually get back
Every capability that costs memory can be switched off on its own, and switching it off releases the memory rather than hiding the feature. Somebody under memory pressure drops a tier instead of dropping the app.
Screen context is off until you turn it on
There is an optional mode that glances at the screen to pick up unusual words. It is off by default. When it is on, a panel names each word it learned and counts itself down, captures are held in memory and never written to disk, and a password manager or an app you excluded is skipped.
Budgets
- Dictation only
- 150 MB
- With vocabulary learning
- 300 MB
- Everything on
- 400 MB
- Ready to dictate, from launch
- 2s
- Feedback after the key comes up
- 150ms
What it needs
Decline the permission that types into other apps and it still transcribes, still hands you the text on the clipboard, and says what you lost. Revoke one later and it notices at the next attempt and offers to open the right settings pane.
Not on sale yet.
Typing Optional will be sold by subscription. The price is not settled, so there is no number here to quote at me later. What is settled is that the people who ask first pay less than the people who arrive at launch, and hear about it before anyone else.
- Distribution
- Direct download, signed and notarized
- Requires
- macOS 26, Apple Silicon
- Account
- None
- On sale
- Not yet
Early Access
One address. You hear the day it goes on sale, at the price this list gets.