Transcription models
From version 1.0.10, the most accurate model, Large v3 Turbo, is part of Pro and your trial. On Free, new recordings use Small (multilingual, 99 languages) once it is on your machine, and your own choice is kept for when Pro is back. An install on Free made before 10 November keeps the model it has until 31 December.
Our recommendation: keep the default. On any Apple-silicon Mac with 8 GB of memory or more, the default model (Large v3 Turbo, quantised) gives the best accuracy for its speed, handles 99 languages, and keeps up with a live call. Setup downloads a smaller model first so you can start recording sooner. The default follows in the background unless you untick it, and new recordings switch to it once it has downloaded. You only need this page if your machine is older or smaller, or you are curious.
Transcription always runs on your own machine. The model is a file you download once, and your recordings and transcripts never go anywhere.
What the choice means for you
- Accuracy. Larger models get names, numbers and accents right more often. The full Large v3 Turbo is slightly more accurate than the default and needs 16 GB of memory.
- Speed. Smaller models finish sooner. On Apple silicon the default keeps up with a live call, so the transcript is written while you talk.
- Disk. The default model is about 574 MB to download. The model setup downloads first, Base (English, 142 MB) or Small (multilingual, 466 MB), stays on disk until you delete it. The smallest model is 75 MB.
If your machine is smaller
- You only need keyword search: Base (English) or Tiny (English) are the fastest and fine for keyword search.
- Meetings in more than one language: stay with the default. The smaller models are weaker outside English.
Change the model at any time under Settings → Transcription. Changing it does not re-transcribe old meetings; choose Transcribe again… from a meeting’s … menu if you want to.
Details
The models
Sizes are for the GGML files Recordist downloads; “quantised” files are the q5_0 versions.
| Model | Download | Languages | Use it when |
|---|---|---|---|
| Tiny (English) | 75 MB | English | Quick drafts, or you only need search |
| Base (English) | 142 MB | English | English calls with clear audio, when speed matters most |
| Small (English) | 466 MB | English | Older Macs where the live transcript matters more than perfection |
| Small (multilingual) | 466 MB | 99 languages | The same, for meetings in other languages |
| Medium (English, quantised) | 539 MB | English | Heavy accents or noisy English audio; slower than the default |
| Large v3 Turbo (quantised), the default | 574 MB | 99 languages | Most Macs |
| Large v3 Turbo (full) | 1.6 GB | 99 languages | Macs with 16 GB of memory or more; slightly better than quantised |
Large v3 Turbo is a version of Whisper large-v3 with fewer decoder layers, so it runs much faster for a small loss in accuracy.
Platform notes
On Apple-silicon Macs, whisper.cpp uses Metal, the Mac’s graphics processor, automatically. There is no setting for the backend. On Windows and Linux, transcription runs on the processor, so it is slower than on a Mac; Recordist on Windows and Linux has the details.
Live transcript
During a meeting Recordist transcribes short chunks so you can read along, and that transcript is the meeting’s transcript. If your machine cannot keep up, the transcript falls behind during the call and catches up after you stop. To rebuild a transcript from the audio, for example with a larger model, choose Transcribe again… from the meeting’s … menu; the new transcript replaces the old one only if the new pass succeeds.
Language
With Auto-detect, Recordist sets the language once two stretches of speech near the start agree, and keeps it for the rest of the recording. Transcribe again… with Auto-detect decides the language for every 30 seconds of audio separately. Pin a language in Settings → Transcription → Language if detection is unreliable, for example when a meeting starts with a lot of names or starts in a different language from the rest.
Voice activity detection
Voice activity detection, in Settings → Transcription, is on by default. The speech model then hears only the parts of the audio where someone speaks, so long silences are skipped. It uses a 1 MB model (Silero VAD) that downloads after your first speech model; without it, the audio is transcribed whole.
Transcribe again…, imports and the recovery of a recording after a crash use this setting and Translate everything to English, as a live recording does.
Managing model files
Settings → Transcription lists the models with their size. A model you have downloaded and are not using has a bin icon next to it for deleting it. Files live in the models folder inside your data folder. Recordist uses only the models in its own list, and only once a file passes its check; any other file you put in the folder is ignored.
Model downloads come over HTTPS from our own copy of the whisper.cpp model files, or from Hugging Face if ours cannot be reached. Each file is checked against a size and SHA-256 built into Recordist before it is used; a file that does not match is deleted and the other source is tried, and if neither gives the right speech model, the app says so. A download that stops continues where it left off. The notes engine and its model are checked as they download, and every download appears in the Privacy Ledger, apart from the few kinds of request listed in The Privacy Ledger, explained.
Whisper is a model released by OpenAI under the MIT licence; whisper.cpp is an open-source project under the MIT licence. “Powered by” statements are factual and imply no endorsement.