Skip to content

Models

Sonari runs two on-device engines: WhisperKit (Apple's on-device Whisper) and NVIDIA Parakeet (via FluidAudio). You choose a model in Settings; it downloads once and is cached locally.

Choosing one

  • Just want it to work? Start with Whisper Base (~145 MB). It's a good everyday default and downloads fast.
  • Want the best accuracy? Download Parakeet v2. It's the most accurate engine Sonari has for English, and still fast, at the cost of a larger download.
ModelEngineSizeNotes
Parakeet v2Parakeet~2.5 GBMost accurate English, very fast
Whisper TinyWhisperKit~78 MBFastest, lowest accuracy
Whisper BaseWhisperKit~145 MBGood everyday default
Whisper SmallWhisperKit~470 MBMore accurate
Whisper Large v3 TurboWhisperKit~632 MBWhisper's most accurate
Whisper Distil Large v3WhisperKit~600 MBNear-large, lighter

Downloading and switching

Models download once and stay cached, so switching between ones you've already downloaded is instant. You can change your model anytime in Settings, and you can even re-transcribe an old recording with a different model to compare.

Where the models are cached is listed under Privacy.