How WhisperType works, and why your voice never leaves the Mac.
Dictation that uploads your voice is a different product from dictation that doesn’t. This page walks through every part of WhisperType — the ways to start it, the six model sizes and what you trade between them, what it does to your words before they land, where the text ends up, and exactly which permissions are needed and why.
01Starting a dictation
A key, or a button that follows you around.
Toggle mode
Press ⌥Space to start, press again to stop and insert. Best for anything longer than a sentence. The shortcut is yours to change, and you can set a separate key for stopping if you would rather not press the same one twice.
Push-to-talk
Hold to speak, release to finish. Best for short bursts where holding is less effort than two presses.
The floating mic button
Turn it on and a round mic button floats above every app, on every Space. Click it to start, click it to stop — it turns red while recording and shows a waveform while transcribing. Drag it anywhere on screen and it stays put between launches. For people who would rather aim and click than remember a shortcut.
Auto-stop and cancelling
Silence ends a recording automatically after 3 seconds by default, adjustable from 1 to 10 or switched off entirely. esc cancels at any point and inserts nothing.
02The HUD and feedback
Knowing what state it is in without looking for a window.
While listening
A floating HUD with a live level meter, so you can see it is actually hearing you — and, with the live preview on (the default), a rough draft of your words as you say them. The draft comes from macOS’s own on-device recogniser and is never what gets inserted; the final text always comes from Whisper.
While transcribing
A spinner, then a soft chime on completion.
At a glance
The menu-bar icon and a status dot show idle, listening or transcribing. The Studio panel spells out what is happening in words — including whether it is downloading model files or loading into memory, each with its own percentage, so a long wait is never an unexplained one.
03Models
What you are trading when you pick a size.
| Model | Size on disk | What it is for |
|---|---|---|
Tiny tiny.en | ~75 MB | Fastest and lightest. Occasional spelling slips. |
Base base.en | ~145 MB | Fast and light. Good for quick notes. |
Small small.en | ~480 MB | The default — the best balance of speed and accuracy. |
Medium medium.en | ~1.5 GB | Very accurate. Uses more memory and time. |
Large Turbo large-v3_turbo | ~3.0 GB | Distilled from Large — nearly as accurate, much faster. Multilingual. |
Large large-v3 | ~3 GB | The flagship, the one used in most published speech-research benchmarks. Multilingual, most accurate, slowest. |
Downloading and removing
Models download on demand, with a progress bar that names the phase it is in. The Models panel shows what is on disk and removes any you no longer want, and Unload hands the memory back without deleting anything.
Why the first load of a big model is slow
CoreML compiles a model for the Neural Engine the first time it loads, then caches the result. Measured on an M4 Mac, that is about 16 seconds for Medium and about 45 seconds for Large Turbo, whose audio encoder now compiles for the GPU instead of the Neural Engine — once. Every load after that is under a second. WhisperType paces its progress bar on the slow case, so it never sits at 95% pretending to be finished.
04What it does to your words
Between the microphone and the text field, in that order.
Silence is trimmed first
The quiet either side of what you said is cut before Whisper ever sees it. This is not cosmetic: fed a long tail of silence, Whisper invents text to fill it — subtitle credits, stray phrases, sometimes in the wrong language. Trimming removes the cause.
Custom vocabulary
Names, jargon and spellings it keeps getting wrong. Your list is handed to Whisper before it transcribes, so a colleague’s name comes out right the first time rather than being corrected afterwards.
Filler removal
“Um”, “uh” and repeated false starts are dropped. On by default, and a single switch away from off.
Text replacements
Your own find-and-replace rules, applied to every transcription. Whole-word and case-insensitive, so “api” becomes “API” without mangling “rapid”.
On-device AI cleanup
On macOS 26 and later, Apple’s built-in on-device model can rewrite the transcription into finished prose — punctuation, paragraphs, the lot. It runs on your Mac, like everything else here; nothing is sent anywhere. Off by default.
Per-app modes
An email in Mail, a one-liner in Slack, literal text in Terminal. WhisperType notices which app was frontmost when you started talking and applies the instruction you wrote for it. Four presets ship (Email, Chat, Code & Terminal, Notes) and you edit them freely; nothing fires until you add one.
Language
Pick any language Whisper supports, or leave it on automatic detection. The multilingual Large models handle accents and code-switching without a cloud service in sight.
05Where the text goes
Two choices, one of them safer for sensitive work.
Auto-paste
A simulated ⌘V into whatever app you are typing in, with your previous clipboard restored a moment later — dictation never quietly overwrites what you had copied.
Clipboard only
Leaves the transcription on the clipboard for you to paste yourself. Slower, but nothing is injected into an app without your action.
06History, stats and the Control Panel
One window, a sidebar, and everything in it.
100-item history
Your last 100 transcriptions, searchable, with timestamps, word counts and one-click copy. Kept on your Mac.
What you saved
Words dictated, sessions, time spent speaking, and an estimate of the typing time you got back. Counted locally, sent nowhere.
Audio input
Record through a headset without changing your Mac’s input device for every other app. Unplug it and WhisperType quietly falls back to the default.
The rest of it
Studio, History, Models, Text & Language, Audio & Input, Shortcuts, Permissions, General, License, More Apps and Feedback — each its own section in the sidebar. General holds launch at login; Feedback sends a bug report or a feature request, with screenshots attached if that is easier than describing it.
07Permissions and privacy
Three permissions, and what each is genuinely for.
Microphone
Required to capture your voice. Audio is transcribed in memory and never written to disk.
Accessibility
Required only to simulate the ⌘V that pastes your text into another app. Without it, dictation still works — the result goes to the clipboard instead.
Speech Recognition
Only for the live preview. Every request is pinned to on-device recognition with no network fallback, so a Mac that cannot do it offline simply shows no preview. Decline it and everything else works as before.
08What it costs
Free forever at 10 minutes a day, or $24.99 once to remove the limit.
Every copy starts with a 14-day trial of the whole app, no account and no card. When the trial ends WhisperType does not go dark: it moves to a free plan of 10 minutes of dictation a day, forever, with every feature still switched on. A one-time $24.99 purchase removes the daily limit on up to three Macs.
09Now go try it.
Free for 14 days, then free every day. No account. Back to WhisperType · Buying & licensing · Support