Offline dictation
Voice to text with the network switched off, on both Mac and Windows.
Most dictation software is a microphone attached to somebody else’s server. That is a reasonable trade for accuracy, and for a lot of writing it is the right one — but it makes the feature unavailable exactly when you tend to need it, and unusable for the material you are least able to send anywhere.
JotMic ships two speech engines and lets you pick per session. The local one runs on your own hardware and makes no network requests to transcribe anything, ever.
What runs on your machine
The local engine is Parakeet — an int8-quantised speech recognition model that lives on your disk after a one-off download and runs from there. On a Mac it uses the bundled build; on Windows it runs through sherpa-onnx. Both are the same model, so the transcript you get does not depend on which machine you are sitting at.
Nothing is uploaded and nothing is queued for later upload. There is no “we transcribe on-device and sync the text” asterisk: on the local engine, the audio and the text both begin and end on your computer. What you dictate is written to a local database you can clear or turn off, described in full on the privacy page.
When to use the hosted engine instead
Honesty about the trade matters more here than a clean pitch. The hosted engine is faster on difficult audio and better at unusual names, because it runs a model far larger than anything sensible to ship to a laptop. If you are dictating in a quiet room with a decent microphone you are unlikely to notice a difference; if you are on a train with an accent the model has not seen much of, you will.
When the hosted engine is on, the audio of the hold is sent and only that — no screenshot, no window contents, no clipboard. It is transcribed, returned and discarded. It is never stored and never used for training.
Who this is actually for
People under an NDA that does not permit third-party processors. Anyone dictating patient notes, client matters, unreleased financials or an unfiled patent. Developers on air-gapped or restricted networks. And, mundanely, anyone who writes on a plane — which is where most people discover that their dictation app is a web service.
The rest of the product does not change
Offline is a switch, not a lesser tier. Automatic punctuation, filler removal, language detection, custom vocabulary and your own modes all work the same way on the local engine, and both engines are included in the free sixty minutes and in every paid plan. On a Mac, voice commands work offline too — reading an application’s menu bar is a local operation and always was.
Questions
Does offline dictation work with no internet at all?
Yes. With the local engine selected, transcription runs entirely on your own machine and needs no connection. Aeroplane mode, a locked-down network or a machine that has never been online all work the same way.
Is offline dictation less accurate than cloud speech to text?
On clear speech the difference is small. The hosted engine pulls ahead on hard audio — heavy accents, background noise, a bad microphone — because it runs a larger model than is practical to ship to a laptop. You can switch between the two at any time.
Does the local engine need a download?
Once. The weights are fetched the first time you select the local engine — roughly 490 MB on macOS and 652 MB on Windows — and then live on your disk. The download resumes if it is interrupted. After that, nothing is fetched to transcribe anything.
Do I still need an account for offline dictation?
Yes, for licensing — the app checks your plan. Your speech is not part of that: no audio, no transcript and no text ever leaves the machine on the local engine.
Try it with the network off. The local engine is on the free tier, so you can prove the claim before paying for it.
60 minutes free · no card