Which Whisper model should I pick?
Recordist transcribes with whisper.cpp on your machine. The model you choose trades disk, memory and speed against accuracy. The default is right for most people.
The short answer
- Apple silicon, 8 GB or more: keep the default,
large-v3-turbo(quantised, about 574 MB to download). Best accuracy, more than 90 languages, and fast enough to keep up with a live call. - Less than 8 GB of memory:
smallorbase. Noticeably less accurate on names and numbers, much lighter. - Speed above all, English only:
base.enortiny.enfinish a one-hour meeting in a couple of minutes on almost anything.
Sizes
| Model | Download | Memory while running | Notes |
|---|---|---|---|
| tiny / tiny.en | 75 MB | ~400 MB | Rough drafts |
| base / base.en | 142 MB | ~500 MB | Good English on old hardware |
| small | 466 MB | ~1 GB | Balanced |
| medium | 1.5 GB | ~2.6 GB | Slower than turbo, not better |
| large-v3-turbo (default) | 574 MB (quantised) | ~1.5 GB | Best quality per second |
Languages
Language is detected automatically. If a meeting mixes languages or detection guesses wrong, pin the language under Settings → Transcription → Language; pinned transcription is also slightly faster. The .en models are English-only.
Changing model
Settings → Transcription → Model. Models download once to the data folder and can be deleted there. Changing the model does not re-transcribe old meetings; use Re-transcribe on a meeting if you want to.
More detail, including the backends on Windows and Linux (early access), is in the transcription models guide.