MacVoice
Swift, AppKit, Apple Speech, MLX-LM
Native macOS voice typing prototype with direct transcription and optional on-device Qwen text cleanup.
Overview
MacVoice is a native menu-bar voice-typing utility for macOS. A keyboard shortcut starts dictation, and recognized text is sent as keyboard events to the currently focused app.
Two modes
- Verbatim: Sends confirmed speech recognition text directly to the typing engine.
- Polished: Buffers phrases and optionally refines punctuation and a limited set of fillers with Qwen2.5-0.5B-Instruct running locally through MLX-LM. If polishing is unavailable or fails validation, the original recognized phrase is used where possible.
Speech recognition uses Apple’s on-device SpeechAnalyzer; speech assets may need to be downloaded by macOS for the selected locale. Polished mode downloads its Python/MLX environment and 4-bit quantized model during explicit setup. Ordinary polishing keeps audio and transcripts on the Mac.
Implementation and status
The AppKit host coordinates audio capture, speech recognition, transcript reconciliation, and the optional local model bridge. A reusable TypingCore component posts Quartz keyboard events rather than using the clipboard. The current distributed build targets Apple Silicon and requires macOS 26 or later; microphone and Accessibility permissions are needed.
MacVoice is a prototype in active development. Transcription quality and text insertion need broader testing, and the repository notes recent reports of repeated or garbled output. App compatibility and recognition quality are not guaranteed.