Skip to content

Transcription engines & models

All voice features - dictation and voice comments and Crit Mode - share one transcription engine, chosen with the Engine picker in Settings > Voice. Everything is transcribed locally on your Mac; no audio or text ever leaves it.

The Engine picker offers three on-device engines:

Engine Download Notes
Apple Speech None Built-in macOS transcription. Good accuracy, works immediately.
WhisperKit 75 MB - 954 MB OpenAI Whisper via CoreML. High accuracy, multiple model sizes.
Parakeet ~1.2 GB NVIDIA Parakeet via CoreML. Fastest inference, excellent accuracy.

WhisperKit offers three model sizes; larger models are more accurate but use more memory.

Model Download size
Fast 75 MB
Balanced 217 MB
Max 954 MB

Parakeet comes in two versions, English and Multilingual, each about 1.2 GB.

WhisperKit and Parakeet models download inside Settings > Voice: click Download next to the model, watch the progress percentage, and use Cancel to abort or Delete to remove a downloaded model later. If a download fails, a Retry button appears.

Voice features keep working while a model downloads: Remarc falls back to Apple Speech until the model is ready. If you switch WhisperKit model sizes, transcription falls back to the previous model while it is still loaded, or to Apple Speech, until the new one finishes.

Two toggles appear in Settings > Voice when WhisperKit or Parakeet is selected:

  • Keep model in memory prevents the model from being unloaded after use. Responses come faster, at the cost of roughly 200-500 MB of RAM depending on the model. Recommended if you have 16 GB or more of RAM.
  • Load model on launch loads the model into memory when Remarc starts so dictation is instant on first use. Available only when keep-in-memory is on.