Transcript quality

The Transcript quality page chooses how accurate the transcript is against how quickly and lightly the Mac makes it. Every choice runs on the Mac.

The modelCopy Anchor Link

The page is three rows, worked out for this Mac. The one in use has a tick; the others are one click away:

• Lighter: a small model. Quick and light on the Mac, with more mistakes.
• Recommended: Large v3 Turbo (quantised). Accurate and quick on most Macs, and the default.
• Most accurate: the full Large v3 Turbo, offered where the Mac has the memory for it. Fewest mistakes; larger and slower.

A choice that is not on the Mac yet shows its download size; one click downloads it and switches to it, with the progress in its row. Exactly one is marked Recommended. A model picked by name under Advanced shows as Your own choice. The choice cannot change while a meeting records.

When to change it: Lighter on an older or busy Mac, or when you only skim transcripts; Most accurate when names and quiet voices matter and the Mac has room.

If it fails: a download that breaks off offers Try again. With no model at all, the page says so and Record stays off until one is downloaded.

The Neural EngineCopy Anchor Link

On Mac computers with Apple silicon, the model's heavy part runs on the Neural Engine, a part of the chip built for this kind of work, instead of the graphics chip. It is not a separate decision: downloading a model downloads its Neural Engine part too, as one size and one progress bar. It is prepared once in the background; until then, and if preparing ever fails, the graphics chip does the work and nothing changes for you.

In our measurement on a ten-minute meeting, the graphics chip did over 90 % less work, at the same accuracy. The Mac stays cooler and the battery lasts longer.

If you installed a model before this existed, the row Make transcription lighter on battery adds its Neural Engine part with Turn on. On an Intel Mac none of this appears.

AdvancedCopy Anchor Link

• Model in use: the model's name, whether it runs on the Neural Engine and its size on disk.
• All models: the full list — Tiny, Base, Small, Small (English), Medium, Medium (quantised), Large v3 Turbo (quantised) and Large v3 Turbo — to download, delete or pick by name.
• Live model: a smaller model just for the words you see during the meeting. The default is Same as the meeting model. It only matters without the Neural Engine.
• Use the GPU instead: runs the whole model on the graphics chip, warmer and heavier on battery. New downloads no longer bring their Neural Engine part, and it offers to delete the ones on the Mac.

Reading the recording again after a live meeting, for a better transcript, is the Final transcript step on After recording.

For the curiousCopy Anchor Link

• The engine is whisper.cpp. On Apple silicon its encoder runs through Core ML on the Neural Engine; the decoder stays on the GPU (Metal).
• The three outcomes map to ggml-small, ggml-large-v3-turbo-q5_0 and ggml-large-v3-turbo. Recommended is computed from the Mac's memory, its chip and whether the Neural Engine is available.
• With a smaller live model, the transcript you keep comes from the meeting model at the end — except with live translation, where the live lines are kept.
• Models are stored in ~/Library/Application Support/Sonavia/Models. Models already in ~/whisper.cpp/models or ~/Library/Application Support/Whisper are used where they are.
• The engine, model and version are written in a comment at the top of each meeting's meeting.md.

What leaves your MacCopy Anchor Link

Only the model downloads, once each, from Hugging Face, pinned to a revision and checked by SHA-256. Audio and text never leave for transcription; there is no cloud speech-to-text in the app. See What leaves your Mac.

RelatedCopy Anchor Link

• Recording — when the transcript is written.
• After recording — the final transcript after a live meeting.
• Transcripts — what the model's work looks like in a meeting.
• Words and terms — names and jargon the model should expect.