VoxBar

VoxBar started as a fork of Handy, an open-source macOS dictation app by cjpais, and grew into a product line of its own: about 50,000 lines added across the Rust tree, from a local LLM post-processing pipeline to a phone companion that pairs over the local network.

The pitch is simple: hold a key, talk, release, and the text lands wherever your cursor is. Speech recognition runs locally, and an on-device language model cleans up the raw transcript in an isolated worker process, with grammar-constrained output and models pinned by hash so nothing can be swapped underneath you. A small state machine keeps the voice model and the polish model from ever sitting in memory at the same time. Local models are available per language, including Arabic, German, Japanese, and Chinese.

The discipline is in the shipping. Nineteen releases went out in five days, up to four a day at the peak, and every one was byte-verified through a 13-gate battery and a 944-function test suite before it went out the door. A small group of daily early users dictates with it today.