LesskeysSpeaker diarization: who said what, offline, on Mac or Windows
Speaker diarization works out who is speaking when, so a transcript shows who said what. Lesskeys does it on your computer with no upload: it separates the voices, labels each line, and you set how many people spoke (2 by default).
macOS 13+ on Apple Silicon · No subscriptionWindows 10 and 11 · No subscription
Last reviewed
What it is
What is speaker diarization?
Transcription asks what was said. Diarization asks who said it.
Speaker diarization splits a transcript into turns: Speaker 1, Speaker 2, or real names once you assign them. Two separate steps produce the words and the labels, and in Lesskeys both run on your computer.
No cloud, no bot
Speaker labels with no cloud and no bot
Many tools upload your audio to label speakers. Lesskeys does it on your computer.
- Nothinguploaded
- No botjoins your call
- Offlineworks with the internet off
Tell the people you record; some places require everyone's consent.
When it is wrong
When speaker labels go wrong
Clear turns come out well. The hard cases do not.
- Two people at once
- Two similar voices
- A word or two
- The wrong speaker count
Too low a count merges two people, too high splits one. Set the right number when you know it, run the file again if the count was wrong, and fix a single mixed-up person by merging or renaming. We publish no accuracy figure for the labels: it depends on your recording.
On a Mac
How to get speaker labels on a Mac
- Choose Transcribe and add a file, or Record Call to capture a call as it happens. macOS asks for Screen Recording permission the first time.
- Say how many people spoke: pick 1 to 4, or enter a number up to 99. The count starts at 2.
- Open Library. Each line sits under a label: rename Speaker 1 to a real name, and merge two labels if one person was split.
A named voice is recognised in later recordings. A voice needs about 15 seconds of speech to be remembered, and voiceprints stay on your Mac.
On Windows
How to get speaker labels on Windows
- Choose Transcribe and add a file, or Record Call to capture a call as it happens. Windows asks for nothing beyond your microphone.
- Say how many people spoke: pick 1 to 4, or enter a custom number. The picker starts at 2.
- Open Library. Each line sits under a label: rename a speaker there, or open Speakers to rename and merge two labels if one person was split.
A named voice is recognised in later recordings. A voice needs about 15 seconds of speech to be remembered, and voiceprints stay on your PC. The first transcription downloads a small set of speaker models; after that it works offline.
Questions
Questions about speaker diarization
What is speaker diarization?
It is working out who is speaking and when, then labelling each part of a transcript by speaker: Speaker 1, Speaker 2, or named people.
Can you do speaker diarization without the cloud?
Yes. Lesskeys does the speaker labelling on your computer, so recordings are not uploaded. It downloads a small set of speaker models once, then works offline.
Does Lesskeys remember speakers across recordings?
Yes. With voiceprints on, which is the default, a speaker you have named can be recognised in later recordings. A voice needs about 15 seconds of speech, and voiceprints stay on your computer.
How accurate is on-device diarization?
It depends on the recording. Clear turns come out well; overlapping speech, similar voices and very short remarks are the hard cases, and a wrong speaker count makes it worse. You can set the count, and you can rename and merge speakers.
More