Compare
Best local AI dictation for Mac: what actually runs where
Local AI dictation means model weights on your disk and inference on your own silicon. Here is which apps do that, which models they use, and where the honest limits are.
Last updated
What changed, and when
Two things made local dictation viable rather than a curiosity. Whisper's weights were published openly, which removed the need to build a recogniser. And Apple Silicon's unified memory architecture let a laptop hold a model and feed it without copying gigabytes between separate CPU and GPU memory.
Neither is a small detail. Before the first, building a dictation app meant licensing a speech engine. Before the second, running one on a laptop meant waiting. See why Apple Silicon suits local speech AI.
The three models a local app is actually running
This is where most comparisons stop too early. A good dictation app runs up to three separate models, and "local" can be true of one and false of another.
| Stage | What it does | Typical model | Usually local? |
|---|---|---|---|
| Voice activity detection | Finds where speech starts and stops | Silero VAD or similar, a few megabytes | Always |
| Speech recognition | Audio to words | Whisper family, Parakeet, Moonshine | Yes in the apps on this page |
| Text rewriting | Tidies the words into finished writing | A small language model, or a cloud API | Often not. Check this one |
Vaitly Voice runs all three locally by default: Silero for voice activity detection, a Whisper-family or Parakeet model for recognition, and a small local language model for the writing step. A cloud provider can be configured for that last stage, and the app labels transforms that use one.
The apps
Vaitly Voice
Ships Whisper-family and Parakeet models and selects for your Mac. The model is downloaded and checksum-verified on first run and lives in the app's own data directory. The writing model is a small local one, which is why the tidying works on a plane. Free is 1,000 words a week with no sign-in.
VoiceInk
Local transcription with the source on GitHub — the only app here whose implementation you can read. Optional cloud enhancement handles transcribed text rather than audio, which the vendor states plainly. Apple Silicon, macOS 14.4 or later. One-off from $25.
Superwhisper
Local and cloud models, with Pro users able to bring their own API keys. Its FAQ notes that offline models run really well on Apple Silicon and that Intel Macs are better served by cloud models. The flexibility is the product; it also means the configuration decides whether you are local.
MacWhisper
Whisper models locally for both file transcription and dictation, with optional connections to cloud AI providers. Pro €64 one-off.
Apple Dictation
Apple's own on-device path, free, and on Apple silicon you can keep typing while dictating. No model choice and no vocabulary, but nothing to install either.
How to evaluate one properly
- Find the model files on disk Application Support, usually. If there is nothing there, recognition is not local whatever the page says.
- Dictate with the network off The basic test, and it settles the recognition stage.
- Now dictate with rewriting turned on, still offline This is the test most people skip. If the tidy-up fails offline, that stage is a cloud call.
- Watch memory while it is warm Activity Monitor. A model held in memory all day is a real cost on a laptop; a good app releases it when idle and reloads on demand.
- Time the first dictation after an idle period A cold model has to load. Sane apps take the hit in the background rather than making you wait.
The honest limits of local
- Model size is bounded by your machine. A laptop will not run what a data centre runs, and on genuinely difficult audio that shows.
- Languages are fewer. Local apps ship what fits. Cloud services ship everything.
- Cold starts exist. The first dictation after an idle period pays a loading cost, unless the app keeps a gigabyte resident all day — which is its own problem.
- Intel Macs are genuinely worse at this. Superwhisper's FAQ says so; it is a property of the hardware, not the software.
- You are choosing a smaller team's roadmap. Local-first dictation is made by small companies. That is a real risk and an honest one to name.
Where these facts came from
Anything stated about another product was read on that product's own site or documentation on the date shown, and is re-checked at least quarterly. If something below has changed, tell us and we will correct it.
- VoiceInk says it "processes all voice transcription locally on your device" and that "your voice data never leaves your Mac," with optional cloud enhancement that handles transcribed text rather than audio. source, checked .
- VoiceInk is open source and published on GitHub for anyone to read, audit or build themselves. source, checked .
- VoiceInk requires an Apple Silicon Mac on macOS 14.4 or later; an iOS app is sold separately. source, checked .
- Superwhisper says it "works offline, so you can transcribe anytime," and its FAQ adds that "Intel Macs work best with Cloud models. Offline models only run really well on Apple Silicon macs." source, checked .
- Superwhisper lets Pro users bring their own API keys and use cloud AI models alongside local ones. source, checked .
- MacWhisper says it uses "local models to transcribe your files" and processes "sensitive content locally without data ever leaving your Mac," with optional connections to cloud AI providers. source, checked .
- Apple's guidance describes continuing to type while dictating on a Mac with Apple silicon. source, checked .
Questions
Which model gives the best local dictation on a Mac?
It depends on what you are optimising. Parakeet models are fast on Apple Silicon and strong on English; Whisper's larger variants are broader across languages and heavier. For live dictation, latency usually matters more than the last point of accuracy, which favours the faster models.
How much RAM does local dictation need?
A dictation-grade speech model is typically a few hundred megabytes to a couple of gigabytes while loaded. On a 16 GB Mac that is unremarkable if the app releases it when idle, and noticeable if it does not. Watch Activity Monitor for an hour before committing.
Is the AI rewriting step local too?
In VV, yes by default — a small local language model does the tidying, which is why it works offline. In many apps it is a cloud API call. This is the stage where a local setup most often leaks, and it is worth checking specifically.
Can I use my own local model?
Superwhisper is the most flexible on this, and VoiceInk's source is open. VV chooses the model for your Mac and does not expose a general model picker in the standard interface — a deliberate narrowing, and a reason to prefer one of the others if you want that control.