Questions
Can speech-to-text work offline? Yes
The recognition is done by a file on your computer. Once it is there, no connection is involved — which is why dictation works on a plane and cloud dictation does not.
Last updated
How it is possible at all
Speech recognition used to require a server because the models were proprietary and large. Two things changed: model weights were published openly, so a model became a file you could download, and Apple Silicon's unified memory made running one on a laptop fast rather than slow.
The result is that the whole recognition step — audio in, words out — happens in your Mac's own memory. There is nothing to send and nowhere to send it. See what local speech-to-text is.
Which apps do it
- macOS's own dictation, on Apple silicon for supported languages. Free, already installed. Check the line under the Dictation toggle in Keyboard settings, because it varies by language.
- VV, VoiceInk, Superwhisper and MacWhisper all run local models.
- Not Wispr Flow, whose privacy page states that transcription always happens in the cloud. Not Google Docs voice typing. Not Microsoft's Dictate, which requires a connection.
The full list with sources is at best offline speech-to-text for Mac.
The one thing that needs a connection
The model download, once, on first run. It is a few hundred megabytes to a couple of gigabytes and it happens before you dictate anything.
This catches people out in exactly one situation: installing an app at the airport and expecting it to work in the air. Install and dictate one sentence at home, and that is the whole preparation. See setting up offline dictation.
Is offline recognition worse?
On clean audio from a decent microphone in a reasonable room: the gap with large cloud models is small enough that whether the app knows your vocabulary matters more.
In a noisy room, or with a strong accent in a second language, or switching between languages mid-sentence: cloud models still lead, because they are larger than anything that fits on a laptop. That is a real difference and it has not gone away.
Language coverage is the other gap. Local apps ship a limited set; cloud services ship everything. If you dictate in a less widely supported language, that decides it before accuracy does.
Prove it on your own machine
- Install and dictate one sentence with the internet on This forces the model download and the permission prompts.
- Turn Wi-Fi off Properly off, not just disconnected. Unplug Ethernet if you use it.
- Dictate into Notes Text appearing means the recognition ran on your Mac.
Two minutes, and it is worth more than any vendor's privacy page — including ours.
Questions
Does offline speech-to-text need to be downloaded first?
Yes, once. The speech model is a few hundred megabytes to a couple of gigabytes and is fetched on first run. After that nothing is needed.
Does Apple's built-in dictation work offline?
On Apple silicon it can, for supported languages. Apple tells you to check the text under the Dictation toggle in Keyboard settings, because it varies by language and region.
Will it work on a plane?
Yes, if the model has already been downloaded. Install and dictate one sentence before you travel. Cabin noise is the bigger obstacle — a headset with a close microphone makes a large difference.
How much disk space does it use?
A few hundred megabytes to a couple of gigabytes, downloaded once and reused indefinitely.
Is offline the same as private?
Nearly. Offline means the app works with no connection. Private means the audio is not transmitted. An app that works offline for recognition is not transmitting audio for recognition — but it may still send something for a licence check or an optional cloud rewrite, which are separate questions.