Choose a Speech Model
The options#
| Model | Pick it when |
|---|---|
| Large v3 Turbo | Default. You want the best accuracy the app offers and your Mac has the headroom. |
| Medium | You want most of that accuracy with a smaller resident footprint. |
| Small | Your Mac is often under load and you dictate mostly ordinary prose. |
| Base | You want the lightest option and can accept more mistakes on technical terms. |
How to decide#
Start with the default. Large v3 Turbo is the recommended model for a reason, and on Apple silicon with Metal acceleration it is fast enough for the hold-and-release loop to feel immediate.
Drop to a lighter model if either of these is true:
- Memory pressure. The selected context stays resident between dictations so the next one starts quickly. That is a deliberate trade of memory for latency, and a lighter model shrinks the memory side of it.
- The Mac is busy. If you dictate while a build or a container stack is running, a smaller model finishes its pass sooner.
What a lighter model costs you#
Accuracy on unusual words, mostly. Common English survives a smaller model well. Product names, library names and domain jargon are where the difference shows.
If that is your vocabulary, two things help: keep the larger model, and add the terms to your personal dictionary.
First-run download#
Whichever model you choose is downloaded once on first run. After that dictation works offline. See Install Phantom Voice.
Next#
- Personal Dictionary — recover accuracy on your own terms
- Local Transcription — what the model actually does
- Back to the documentation home