STT engines

STT (Speech-to-Text) is the engine that turns speech into text. Whispree offers three of them, and you can switch anytime from Settings → STT or the home dashboard.

Engine comparison

Engine	Location	Model	Requirements	Notes
WhisperKit (default)	Local (CoreML + Neural Engine)	whisper-large-v3-turbo (~1.5 GB)	Model download on first use	99 languages, injects domain words as recognition hints. Works without internet.
Groq Cloud	Cloud	whisper-large-v3-turbo (server)	Groq API key	Very fast (~200ms). Converts audio to WAV and uploads.
MLX Audio	Local (Python worker)	Qwen3-ASR-1.7B-8bit (~1.0 GB)	uv installed	Strong in Korean/Chinese/Japanese/English. First load (cold start) ~1 min. Swappable to any mlx-audio model.

Tip

For the simplest start, leave it on WhisperKit (local, no setup needed). If you want a faster cloud option, get a key from console.groq.com and use Groq. The Groq key is shared across both STT and correction.

Language

Pick the recognition language in Settings: auto-detect or ko · en · ja · zh · es · fr · de · pt.

• Auto-detect is convenient but can be less accurate, so a warning is shown. If you mostly speak one language, pinning that language is more accurate.
• WhisperKit has a correction step in auto-detect: if it mistakenly detects a language other than Korean/English/Japanese/Chinese, it retries once in Korean.

The selected language is also used to pick the AI correction prompt (Korean↔English code-switching, etc.). See AI correction.

Silence auto-skip (VAD)

Silence auto-skip (on by default) in Settings → STT uses voice activity detection (VAD) to remove speechless segments before transcription. It applies to every STT engine, making transcription faster and cleaner, and shows a “skipping silence” indicator during recording.

Domain word recognition

WhisperKit and MLX Audio take the domain words you register in the dictionary as recognition hints, improving dictation accuracy for technical terms and proper nouns.