Choose a local speech model for Soufflé

Kyutai streams text directly. Whisper and Parakeet decode audio windows while recording, so previews may change as more context arrives; stopping finalizes the remaining audio. Apple Speech provides progressive recognition on supported systems. All use the same Me/Them source distinction for meetings.

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, Parakeet and available Apple Speech. Whisper and Parakeet show revisable live previews and finalize at stop; all can keep Me/Them sources separate in meetings.
  3. Select the engine and language in Settings → Transcription and finish downloading before going offline. Apple Speech requires supported macOS 26 hardware and language assets; macOS manages those assets.

Speech models

Speech recognition runs locally. Downloadable checkpoints come from Hugging Face; Apple Speech uses language assets installed and managed by macOS. Choose an engine in Settings → Transcription when no recording is active.

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. Revisable live previews, then final text when you stop; separate Me/Them in meetings.
Parakeet TDT 0.6B v325 languages with punctuation and capitalization, ~670 MB int8, CPU inference via ONNX Runtime. Revisable live previews, then final text on stop; separate Me/Them in meetings.
Apple SpeechOptional on supported macOS 26 Macs and locales. On-device recognition with macOS-managed language assets, progressive text and separate microphone/system-audio results. Availability and installation guidance appear in Settings → Transcription.

Kyutai streams text directly. Whisper and Parakeet decode audio windows while recording, so previews may change as more context arrives; stopping finalizes the remaining audio. Apple Speech provides progressive recognition on supported systems. All use the same Me/Them source distinction for meetings.

Apple Speech transcribes audio. It is separate from Apple Intelligence, which is used for optional dictation polish and experimental meeting summaries.

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.