Choose a local speech model for Soufflé

Every speech model runs on your Mac. Choose by language first, then by whether you want text during recording or once it ends. The model download is separate from the app and is kept for later use.

Updated

Start here

  1. For streaming dictation in French and English, start with Kyutai STT 1B. It is also suitable for Me/Them separation in meetings.
  2. For other languages, compare Whisper and Parakeet. Both transcribe once recording stops, and neither separates the microphone from the other meeting voices.
  3. Choose the model in Settings → Transcription and download it before going offline. Check the disk requirements below; unused models can be removed in Settings → System.

Speech models

All models run locally and are downloaded on first use from HuggingFace. Switch at any time in Settings → Transcription.

Kyutai STT 1BDefault. French and English, ~2.4 GB, Metal GPU via Candle. Streams text while you speak.
Kyutai STT 2.6BEnglish only, higher quality, ~5.6 GB. Streams text while you speak. It has no pause detector, so long speech without a quiet gap can lose about 2.5 seconds every 28 seconds.
Whisper Large V3 TurboMultilingual, ~1.6 GB, Metal via whisper.cpp. Transcribes once you stop.
Parakeet TDT 0.6B v325 languages with punctuation and capitalization, ~670 MB int8, fast CPU inference via ONNX Runtime. Transcribes once you stop.

The two Kyutai models are the streaming ones: text appears while you are still talking. Whisper and Parakeet transcribe the whole recording once you stop, which is fine for a meeting you read afterwards but changes how dictation feels.

A loaded model is unloaded after an hour of inactivity to give the RAM back, and reloaded on the next recording. Change or disable that in Settings → Transcription.