Chatterbox 1.7.2 for Linux

Live local speech-to-text captions for the VRChat chatbox. Linux x86_64, one self-contained file.

New in 1.7.2: the binary is never changed under a running process. A single-file app reads every assembly it has not used yet from its own executable, so replacing or deleting that file while the app ran could make a later load fail. --uninstall from the installed copy now removes the app folder as its very last act, and an in-app update is swapped in by a small helper only after Chatterbox has exited, which then starts the new version. New --update switch (the Updates screen from a terminal); --uninstall --purge also removes the unpacked native libraries under ~/.net/.

Also a stability audit of the whole app: a recognition pass that throws stops the session at once with "recognition failed", and Stop is quick; a CPU without AVX2/FMA gets the bundled no-AVX Whisper build, so both engines run there; the crash marker covers the whole run, so a crash during captions is recorded and the next start does not walk back into it; the Translate tab says when the Vulkan pack is not doing the work, and why; the microphone is chosen by the name you see; settings survive a slow-mounting home folder in full; an in-app restart of a Steam-launched instance still exits with VRChat; model, engine and GPU-pack downloads resume after a network failure. tools/scenarios.sh and tools/soak.sh run the checklist's no-VRChat scenarios and a long stability soak.

This binary was run headless in WSL2 Fedora 44 with tools/smoke.sh and tools/scenarios.sh; it has not been run on a real Linux desktop with VRChat yet. Reports welcome.

Download Chatterbox-1.7.2-linux-x64, then chmod +x it and run it. The in-app updater (Settings → Updates) also offers this release to earlier installs.

SHA-256: E4703F377484B8C0DE1A66D0ED5D16110A7EEACD22A8926D031A591521D8AAA6

Chatterbox 1.7.1 for Linux

Live local speech-to-text captions for the VRChat chatbox. Linux x86_64, one self-contained file.

New in 1.7.1 — translation. Your captions can be translated on your own machine before they reach the chatbox: Tencent's Hy-MT2 1.8B model (Apache-2.0, 1.1 GB) through llama.cpp, fetched on demand from the Models screen as the model plus a 36 MB CPU engine pack; an optional 21 MB Vulkan pack runs it on any GPU (NVIDIA, AMD or Intel) through your distribution's Vulkan loader. The new Translate tab holds the switch, the language and "show the original too", and shows the last translated sentence. The recognition engine and Whisper model selectors moved from Settings to Models → In use. Nothing is sent anywhere — the model and the runtime both live on your machine. Details: README "Translation" and docs/TRANSLATION_BENCH-2026-10-08.md.

Checked in WSL2 Fedora 44 (CPU path); not yet run on a real Linux desktop — reports welcome.

Download Chatterbox-1.7.1-linux-x64, then chmod +x it and run it. The in-app updater (Settings → Updates) also offers this release to 1.6.0 installs.

SHA-256: 602B501FA35DC0338936824315092BF66DF6638ABBBAECD57BD4A008AEF24D9F

Chatterbox 1.6.0 for Linux

Live local speech-to-text captions for the VRChat chatbox. Linux x64, one self-contained file (chmod +x after download). New: Settings → Updates checks this page for newer versions and installs them in place (off at startup unless you turn it on). SHA-256: 2FF78F4868893F6BAD4065F095A9FD3E036129088749FD0310A2D036D6B3A0AC

Chatterbox 1.5.4 for Linux

Live local speech-to-text captions for the VRChat chatbox. Linux x64, one self-contained file (chmod +x after download). SHA-256: FD71954F56D614BD4E0E375682446A3EAF2CAB8D595F5DF6A82A45045D7A6521