Listen To
All system audio, which is the default and needs nothing set. The same menu lists what is playing right now if you would rather narrow it to one app, and that choice sticks until you change it.
Subtitles · Setup guide
Set it up once,
then forget it is there.
Installing takes a minute. The first launch takes longer, because the speech model downloads before anything can be captioned. One step in the middle can go wrong silently, and this page exists mostly so you can spot it in ten seconds rather than an hour.
Before you start
- macOS 14.2 or later, on Apple silicon. The whole app is built on Core Audio process taps, which is the API that arrived in 14.2. An Intel Mac cannot run it at all, and no update will change that.
- About 600 MB of disk, once. The speech model is not in the download. It is fetched on first run and then kept: 583 MB for the Latin-script pack, 633 MB for the one that includes Chinese and Japanese.
- A few minutes of internet, once. That download is the only network call the app makes in its life. Everything after it works offline.
First launch
Drag it to Applications and open it. There is no Dock icon and no window to arrange: it lives in the menu bar. What you get first is a short welcome that shows the overlay doing its three jobs.
Behind it, the model is downloading. The menu bar icon carries a pulsing dot while that runs, and the menu itself shows a percentage and a file count. First launch takes minutes; every one after it takes seconds.
The step that fails quietly
macOS asks for permission to record system audio. If you miss that prompt, or dismiss it, nothing will tell you. Core Audio keeps returning success, the tap reports itself running, buffers arrive at the right size and cadence, and every single sample is zero. There is no error anywhere in the stack, so the app looks installed and working and simply never captions anything.
If you see no captions and no error, this is almost always why. The menu has Check Audio Permission…, which opens the right pane of System Settings; Subtitles should be ticked under Privacy & Security ▸ Audio Recording. There is no API that reports whether the grant exists, so the app does not pretend to know.
Where it lives
Language / Models
Auto-detect leads, and covers nine languages. English has its own models with a speed and accuracy trade-off; the other languages come from one multilingual checkpoint. Moving within a pack is instant, crossing between packs is another download.
The switches worth knowing
Skip Non-Speech keeps music from becoming invented lyrics. New Box On Speaker Change is off by default and worth turning on for calls. Fade Away Under Pointer stops the overlay hiding what is beneath it.
Settings
Box size, opacity, how many lines it keeps and how long the history lives, all previewed on the real overlay rather than in a mockup. The model cache can be cleared from the same window.
The four keys
The overlay takes no clicks, so you work straight through it. These are the only things it wants from your keyboard.
- Hold ⇧ and drag to move the overlay. Where you put it is remembered.
- Hold ⌥ to bring the last few captions back and scroll through them.
- ⌥⌘S pauses and resumes captioning.
- Point at it and it fades, so you can read what is underneath without moving anything.
When something is wrong
- Nothing is being captioned and there is no error.
- The audio permission, nine times out of ten. Open Check Audio Permission… in the menu and make sure Subtitles is ticked under Audio Recording. The failure is silent by design of the API rather than by choice: a denied grant and a quiet machine are identical from inside the app.
- Captions are appearing but the words are wrong.
- Check which model is loaded. Auto-detect is the most convenient and not the most accurate; if you are only listening in English, picking an English model helps, and picking a slower one helps more. Names, jargon and acronyms are where any recogniser slips, and no setting fixes those.
- Music is turning into lyrics.
- Skip Non-Speech (VAD) is switched off. Turn it back on and a voice detector runs ahead of the recogniser, so backing tracks never reach it. It is on by default.
- The overlay is covering something I need to see.
- Hold ⇧ and drag it somewhere else, or leave the pointer over it and it dissolves. How far that reach extends and how much it fades are both settings. If it has ended up somewhere unreachable, Reset Overlay Position in the menu puts it back.
- It stopped captioning after the Mac slept.
- Toggle captions off and on from the menu, which re-establishes the tap. If the app is otherwise responsive, that is faster than quitting it.
- I want the disk space back.
- Settings has Clear Model Cache, which removes the downloaded weights and reports how much it recovered. The next launch downloads again, so do it when you are done rather than between sessions.
The same overlay, elsewhere
Learning a language
Subtitles in the language being spoken, not translated into yours.
vs Apple Live Captions
macOS has this built in and free. Where it stops being enough.
Zoom
Captions without the host switching anything on.
Microsoft Teams
Captions that do not depend on what your tenant's admin allows.
Google Meet
On-device captions, and no Workspace tier to be on.
Slack huddles
Captions with no workspace app to get approved.
Lectures
Captions live, rather than once the recording is processed.
Podcasts
Captions for the shows that never had a transcript.
vs Granola and notes apps
Complementary, not alternatives. Notes after, captions during.
Set up in a minute, then out of the way.
Launch week: $5 instead of $9 until 22 August · 14-day refund
14-day refund, no questions asked
One-time $9, every future update included ·
macOS 14.2 or later · Apple Silicon
Unlimited live captions, for as long as you own the Mac. No minutes to
buy, no account, no subscription.