Repository navigation
Live local speech-to-text captions for the VRChat chatbox. Windows 10/11 x64.
New in 1.7.2
- In-app updates never change the exe under a running process any more. The download is staged beside the running exe, Chatterbox exits, and the staged exe swaps the files once the old process is gone and starts the new version. (You receive 1.7.2 itself through the old path one last time; later updates use the new one.)
- Being ended from outside is not a crash. The crash marker now covers the whole run, and the next start asks Windows whether it recorded a crash or hang for it: if so, captions wait for Start and the record is copied to error.log; a run that was simply ended (Task Manager, or a companion app that stops Chatterbox when the game closes) auto-starts captions as usual. Sign-out and shutdown count as clean exits.
- Stability: when recognition cannot keep up, audio beyond one pass plus ten seconds is skipped instead of lagging further; a recognition pass that throws stops the session with "recognition failed"; Stop is quick and never frees the engine under a running pass; Parakeet gets at most 20 s per pass.
- A model, engine or GPU-pack download that fails on the network keeps its partial file, and the retry continues where it stopped.
- A CPU without AVX2/FMA gets the bundled no-AVX Whisper build, so both engines run there.
- The Translate tab says when translation runs on the CPU although the Vulkan pack is installed, and why. The microphone is chosen by the name you see. A settings save that fails says so.
Run on Windows 11 before release: the update from an installed 1.7.1, the new update path, a kill during captions (CPU, and CUDA with translation), a real crash record, a first-run download and a resumed one.
SHA-256: D7BE5D52BE0154DA6B4848CBA0D4C241E0A6E9A4919AE5A0624E45AA1FE1D6F9