How local transcription works on a Mac or a PC
Local transcription turns speech into text with a speech model that is stored on your computer and run by its own processor, so no server and no internet connection are involved. Audio goes in, five steps run on the computer, and text comes out.
macOS 13+ on Apple Silicon · No subscriptionWindows 10 and 11 · No subscription
Last reviewed
Five steps
What happens between speaking and the text?
- Listen. Audio comes from your microphone, a file, or the sound your computer plays.
- Tidy up. The app finds the speech, evens out the volume and cuts the recording into pieces.
- Write it down. The app turns each piece into text, right on the computer.
- Tell speakers apart (optional). Voices are grouped by speaker. You set how many people spoke (2 by default).You say how many people spoke (1 to 4, or a custom number) and Lesskeys groups the voices.
- Put it together. The pieces become one transcript, with speaker labels if you asked. Every step ran on the computer.
Speaker labels: speaker diarization.
On your computer
What is running on my computer?
Everything it needs sits on your computer.
A small starter is inside the app, so dictation works from the first launch; bigger ones download once and stay. It needs macOS 13 or later, Apple Silicon (M1 or newer)Windows 10 (version 1903 or later) or Windows 11.
Internet
Does it ever need the internet?
Only to download things once.
A bigger, more accurate download if you choose one, and a small extra one the first time you label speakers. After that, no connection is needed. What is on-device transcription compares local and cloud; a test you can run.
On a Mac
Is it fast enough on a Mac?
In our test on one Mac, an hour of speech took about five minutes with the more accurate option and about two minutes with a smaller one.
An older Mac is slower. Dictation writes the text after you stop talking: a short wait, not live captions. The work runs on the Mac's own Apple Silicon chip.
Source: our own measurement, in Which Whisper model to run on a Mac.
On Windows
What does the PC do?
A graphics card with Vulkan support does the heavy lifting; without one the processor does it.
On a 64-bit Intel or AMD PC, Lesskeys uses a graphics card that supports Vulkan when it finds one, and falls back to the processor when it does not. The ARM64 build runs on the processor. Dictation writes the text after you stop talking: a short wait, not live captions. We have not measured how long an hour of speech takes on Windows, so there is no figure for it.
Questions
Questions about local transcription
How does offline transcription work?
Software stored on your computer turns the audio into text on your machine, with no server involved, so it works with the internet off.
Is local transcription slow?
No. In our test on one Apple Silicon Mac, an hour of speech took about five minutes with the more accurate option; an older Mac is slower. We have not measured Windows.
How does it label different speakers without the cloud?
Separate tools on your computer compare the voices across the recording and group the segments by speaker. You can set how many people spoke, and nothing is sent to a server.
More