Local Whisper stores macOS service settings in ~/.whisper/config.toml. Change most settings from the app or with wh config. Edit the TOML directly when you need exact control. Qwen3-ASR defaults to the higher-quality mlx-community/Qwen3-ASR-1.7B-bf16 checkpoint. Set qwen3_asr.model to mlx-community/Qwen3-ASR-0.6B-bf16 for lower memory use and latency at lower quality. These are community-maintained MLX conversions used through a community-maintained runtime, not Qwen’s official PyTorch stack. With qwen3_asr.use_vocabulary = true, enabled Vocabulary replacement rules become local Qwen context and hotwords. Both catalog models support the capability. Empty/deletion replacements are omitted, the full spoken and preferred forms are retained, and the generated prompt is capped at 4,096 characters. Unsupported engines and unvalidated custom model IDs receive no Vocabulary context.

Full reference

Source of truth

The Python schema and defaults live in src/whisper_voice/config/schema.py. When a setting changes, update this page and the schema together. Hold-to-record always pastes the transcription at the active cursor. auto_paste = true makes double-tap dictation use the same paste-at-cursor behavior.