You can, and there's no model to download: macOS has shipped the engine since version 26. The odd part isn't that it works — it's that almost no dictation app uses it.
01 · The problem
They send the audio to a server, transcribe it there and send the text back. They do it because until recently that was the only way to get good accuracy, and because owning a server is what lets you charge a subscription.
02 · The other way
Since macOS 26, Apple ships SpeechAnalyzer, a speech recognition engine that runs on the device itself and covers 30 locales.
VoiceFlow is built on it. That's why the download is 6.5 MB and why it fetches no model at all: what it needs already came with the system.
The test takes ten seconds: turn off the Wi-Fi and keep dictating.
03 · The detail
Being honest about this matters more than the promise, because it's what you find out on day two.
The rule-based cleanup — filler words, punctuation, spoken corrections — uses no language model at all: they're deterministic rules that take under 10 ms per dictation. That figure comes from the project's own test bench, which runs on every change across 62 cases.
| Feature | With no internet |
|---|---|
| Transcribing what you say | Works |
| Removing filler words and punctuating | Works |
| Resolving spoken corrections | Works |
| Your own vocabulary | Works |
| Typing where your cursor is | Works |
| Rewriting the whole text | Works, with Apple Intelligence |
| Translating a selection | Downloads the language once |
| Checking for updates | Needs a connection |
04 · The obvious question
It also works offline, and it uses the same engine. The difference isn't in the transcribing: it's in what happens next.
The whole story is in the comparison with macOS dictation, with the same audio run through both.
Try it
With 14 days of everything in Pro. No account, no email and no form: you download it and that's it.
Version 1.4.0 · 6.5 MB · Intel and Apple Silicon · requires macOS 26