Settings
Transcription models
Pick the on-device speech model that fits your Mac
Dictato runs two model families, both entirely on-device. Pick one in Settings → Model — unselected models download on demand.
Parakeet (recommended)
NVIDIA's Parakeet family, running through Apple Neural Engine. Fastest on modern Apple Silicon.
| Model | Size | Best for |
|---|---|---|
| Parakeet V3 | ~600 MB | Multilingual (25 languages), ~210× real-time on M4 |
| Parakeet V2 | ~400 MB | English only, highest recall |
WhisperKit
OpenAI's Whisper models compiled for CoreML. Slower than Parakeet but more widely supported.
| Model | Size | Best for |
|---|---|---|
| Base | ~150 MB | Lightweight, good for most languages |
| Distil Large V3 | ~600 MB | 6× faster than Large V3, within 1% accuracy |
| Medium | ~1.5 GB | Higher accuracy across languages |
| Medium (English) | ~1.5 GB | Higher accuracy for English only |
Which should I pick?
- English only, fast machine: Parakeet V2
- Multilingual: Parakeet V3
- Older Mac or wide language support: WhisperKit Base or Medium
- Best accuracy, don't mind size: WhisperKit Medium or Distil Large V3
Info
Models download the first time you select them. Parakeet downloads lazily on first use; WhisperKit downloads up front so you can cancel and retry.