The speech engines, AI processing and drafting, the terminal Agent, personalization, and the questions people ask.
Recognition runs entirely on your Mac. Start on the bundled one, switch anytime in Settings, no re-install.
| engine | size | best at | availability |
|---|---|---|---|
| SenseVoice | 228 MB | Fastest (~0.06s). Chinese-first; speaks zh / en / ja / ko / yue mixed in one sentence. | bundled · default |
| FireRed | ~740 MB | Steady on long recordings; handles 20+ dialects. | free download |
| Qwen3-ASR | ~1.8 GB | Best zh/en mixing. 30 languages + 22 Chinese accents. Steady memory; your corrections list doubles as built-in hotwords. | free download |
No API key required. Want it hassle-free? Run the local model. Want more speed and quality? Plug in your own API.
Runs on your Mac. Offline and free. Loads only for the few seconds it's cleaning up, so idle memory is zero.
Routes by intent: a quick tidy takes ~1.7s; "write me an email" thinks for ~10s.
Plug in any OpenAI-compatible endpoint for top speed and quality.
Your key lives in the macOS Keychain, never uploaded and never written to a log.
The backend tries in order: BYOK → local Qwen → regex fallback. If one fails it steps down, so it never loses your words.
Great defaults out of the box. Or keep tuning until it fits how you talk and work.