StudioDeveloper releases Vocs: fully local macOS dictation with whisper.cpp, open source
Original title: Fully local dictation on macOS with whisper.cpp (large-v3-turbo): streaming chunks, paste anywhere, open source
A developer released Vocs, a MIT-licensed macOS dictation app that runs Whisper (large-v3-turbo) fully on-device and pastes text into any app. Audio is chunked at natural silence boundaries and decoded while speaking, with a window sized per chunk making short fragments about 8x faster than the full 30-second window. Hallucinations are handled by stripping sound-effect tags like [BLANK_AUDIO], detecting and retrying repetition loops, plus a deterministic Rust post-processing pass for punctuation, casing, fillers, spoken commands and a custom dictionary. An optional ~120 MB on-device grammar model applies only safe word-level fixes. No network calls in the transcription path; requires macOS 13+, built with Tauri and Rust, available on GitHub.
The developer shares implementation details on streaming chunked decoding and hallucination handling, useful reference for anyone building local speech input.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.