Voice Models
Manage local Whisper models — download, delete, and pick per Profile.
The Voice Models screen manages the local Whisper models that do on-device speech-to-text. It's pure model management: download, see install state and size, and delete. There's no global "active model" — each Profile picks its own.
The model library
Models range from tiny (fast, small, less accurate) to large-v3 (slow, large, most accurate). Bigger models need more disk and RAM and take longer to transcribe. For most dictation a small or medium model is a good balance; pick a larger one if accuracy matters more than speed.
Some models are language-specific — Kotoba Whisper, for example, is Japanese-only and can be selected only when the Profile's recognition language is Japanese.
Each row shows the download size and whether it's installed, with a button to download or delete. The default model (Whisper Turbo) can't be deleted — it's the fallback everything else relies on.
Downloading
You can download a model from the Voice Models screen, or directly from a Profile's voice-model dropdown (undownloaded models show a download button). A progress overlay appears and can be cancelled and retried.
Idle unloading
To free memory, a loaded voice model is unloaded after an idle period (set by Voice model active duration in Settings). The next dictation reloads it automatically — you don't need to do anything.
The model loads lazily when a dictation starts, not at app launch. If the active Profile's model isn't downloaded yet, recording won't start — you'll get a prompt to download it first.
On-device with Apple Speech
Besides Whisper, a Profile can use Apple Speech — Apple's built-in on-device speech-to-text. It needs no API key and keeps audio on your Mac. It requires macOS 26 or later; on earlier macOS the Profile falls back to local Whisper. Like Whisper models, its per-language assets download from this screen and are selectable only when the Profile's recognition language matches.
Cloud speech-to-text
Prefer cloud accuracy? A Profile can use OpenAI or Groq speech-to-text instead of a local model. It uses your own API key, which you can save either in this screen's Cloud section or in LLM Models — both edit the same key (a provider's one key covers both its speech-to-text and LLM cleanup).