Two models, both local
Whisper Small — the everyday model. About 487 MB. Quick to download, quick to answer, and light enough to leave running all day. It gets unusual names and technical jargon wrong more often than the big one.
Whisper Large v3 Turbo — the accurate one. About 1.6 GB, a little over three times the download, and it holds more memory open while you dictate. Noticeably better on names, accents and technical words, which are exactly the cases people notice.
Both run entirely on your Mac. Neither uploads audio, on either plan.
On macOS 26 and later there is a third row, and it is not a Whisper model: Apple’s own. It is further down this page, with what it costs you.
Which one you get
YapScribe reads how much memory this Mac has and pre-selects the model that suits it, with the reason written out — “Your Mac has 32 GB of memory, which is enough to run the accurate model comfortably”. You can overrule the suggestion; it is a default, not a decision.
The one exception is a model that genuinely will not work. Whisper Large v3 Turbo needs at least 16 GB of memory, and on a Mac with less it is shown greyed out with both numbers on screen — “Needs 16 GB of memory — this Mac has 8 GB” — rather than hidden. You can see the rule instead of being told the answer.
Where the model comes from
From Argmax’s WhisperKit model repository on Hugging Face, once, in the background. YapScribe tells you the size before it starts and lets you know when the model is ready, so you can carry on setting up in the meantime. After that the model lives on your Mac and the download does not happen again.
The Neural Engine
The model runs on the Apple Neural Engine through Core ML. That is why YapScribe is Apple Silicon only and why there is no Intel build: an Intel Mac has no Neural Engine, and running a speech model on the CPU instead would saturate a core, spin the fans and eat the battery for the same words.
If you look at Activity Monitor during a dictation you will see one busy CPU core. That is what healthy decoding looks like — the loop that asks for the next word runs on the CPU while the model’s own arithmetic happens on the Neural Engine.
Changing your model
You choose during setup, and at that moment nothing has been loaded yet, so your choice is simply the one that is used.
To change it later, open the YapScribe menu in your menu bar and choose Change Speech Model…. It opens setup’s model card on its own, with the model you use now already picked — the same rows, with the same reasons written out — and you pick a different one. A model that is not on your Mac yet downloads in the background, the way it did the first time, and the menu item stays grey until that download has finished.
A Whisper model you switch to after YapScribe has already loaded one this session takes effect at the next launch — and the app says so rather than pretending otherwise: “YapScribe will start using it the next time you launch it — this session already has another model loaded.” The first time a model is used on a Mac, Core ML compiles it for that machine, which is not instant; throwing a compiled model away mid-session would make your next dictation pay that cost all over again.
Apple Speech, on macOS 26 and later, is the exception: there is nothing for YapScribe to compile, so turning it on or off takes effect straight away — once macOS has its model, if it had to fetch one (see below).
There is no model manager in Settings yet. The menu item is the way to change.
Apple Speech, on macOS 26 and later
macOS 26 brought a speech model of Apple’s own, and on a Mac that has it YapScribe offers it as a third row. It is the same summary the app’s model card shows you: the speech model that is already inside macOS. No download from YapScribe.
It is free and quick — it is part of the system, so there is no download and no wait at launch. It is also never the suggestion: the recommendation is still a Whisper model chosen by how much memory your Mac has, and this row sits under them for you to pick or ignore.
It listens in one language, and that language is your Mac’s
This is the part to read before you choose it. Apple’s engine has no language detection: it has to be told which language to run, so YapScribe tells it the language your Mac is set to, and it transcribes in that language whatever was actually said.
- It cannot tell languages apart the way Whisper does, so a dictation in another language — Turkish, say — comes out wrong rather than switching.
- Whisper takes the dictation instead only when your Mac’s language is one Apple does not cover. In that case the row cannot be picked at all, and it says so: “Apple’s speech model does not cover the language this Mac is set to. Whisper does.”
Whisper is the engine that listens and decides, which is why it stays the default. If you dictate in more than one language, that is the reason to leave this row alone.
What it costs you
- No Turkish, and it gets technical jargon wrong about three times as often as either Whisper model — which is precisely the case YapScribe exists for.
- It cannot be nudged with your screen’s words the way Whisper can, though your saved words are still applied afterwards, by the same pass that runs on everything else.
- YapScribe still downloads a Whisper model to fall back on, and uses it whenever Apple’s cannot take a dictation. Choosing this row does not save you the Whisper download.
The languages, and where to read the current list
Apple adds languages between releases, so YapScribe counts them on your Mac rather than printing a number that would go stale — and this page does the same. The app’s model card names the language it will listen in and how many Apple covers, on the Mac you are reading it on. That is the current list; anything written here would be last year’s.
Nothing of ours is downloaded, and macOS says where it has got to
Picking this row is what asks macOS to fetch its model, if the Mac does not already have it — YapScribe downloads nothing here, and nothing happens at all on a Mac whose owner never picks the row. While macOS is fetching, the row carries its own line — “Installing Apple’s speech model — 40%. YapScribe uses Whisper until it lands.” — so the wait is on screen rather than a surprise at your first dictation.
Parakeet is not offered
Parakeet is a fast English speech model that other dictation apps offer, and YapScribe does not have it. We benchmarked it: on English it matches Whisper Large v3 Turbo at several times the speed, and on Turkish it produced confident transliterated nonsense — a 98% word error rate — because it has no Turkish at all.
A model that is excellent in one language and useless in another cannot be the default in a multilingual dictation app, and it is not shipped as an option today either. If that changes, it will be as a clearly labelled “Fast English” choice, never as the default.
On macOS 26 and later the Apple Speech row above is that kind of option already — fast, narrower than Whisper, never the default, and with the same gap, no Turkish. On macOS 14 to 25 there is no fast-English choice, and Whisper is the whole list.