Skip to content

Releases: kaorii-ako/VOX

Vox v0.2.2

Choose a tag to compare

@github-actions github-actions released this 17 Jul 06:12

What's fixed

  • LLM rewrite pass no longer answers your dictation. Dictated questions ("do you want to go with a race") were treated as questions addressed to the model and got answered instead of transcribed. The transcript is now clearly framed as content to clean up, never a request.
  • First macOS release. Fixed the macOS build (whisper.cpp needs a 10.15+ deployment target). Builds are unsigned for now — right-click > Open on first launch.
  • Windows ARM64 builds are dropped until whisper.cpp supports the toolchain.

SHA256SUMS attached.

Vox 0.2.1 — push-to-talk fixed

Choose a tag to compare

@kaorii-ako kaorii-ako released this 16 Jul 04:08

Vox 0.2.1 — push-to-talk dictation, now actually working

Hold a key, speak, release — cleaned-up text pastes at your cursor. Free local transcription (embedded whisper.cpp), optional LLM cleanup (Ollama or any BYOK provider), no servers, no accounts, no telemetry.

Fixed

  • Push-to-talk hotkey never registered. The UI offers bare modifier keys (left/right Ctrl, Alt, Shift) as the dictation binding — including the default — but the OS global-shortcut plugin rejects bare modifiers, so dictation silently never started. Bare-modifier bindings now use physical key-state polling (~25 ms edge detection); key combinations (Ctrl+Space, F9, …) still use the OS shortcut path. Hotkey registration failures are now reported instead of swallowed.
  • whisper.cpp/ggml debug output no longer sprays raw stderr; it is routed through the app's logger.
  • The System panel no longer claims X11 paste requires xdotool — keyboard simulation works natively on X11; xdotool is only a fallback.

Added

  • examples/e2e.rs smoke test: mic capture, local Whisper transcription, Ollama rewrite, real clipboard + Ctrl+V paste, and hotkey simulation against a running app.

Verified end-to-end (Linux X11)

Hold ControlLeft → record → whisper base.en transcribe → llama3.2 rewrite → paste into a live window → previous clipboard restored.

Downloads

Linux x86_64 packages are attached (.deb and .rpm), built and boot-tested on this release. No AppImage this time (linuxdeploy failed in the build environment) and no macOS/Windows installers — build those from source with pnpm tauri build until CI release builds exist.

Known limitations

  • Bare-modifier hotkeys need X11 (or XWayland). On pure Wayland, key-state polling is unavailable — pick a key combination (e.g. Ctrl+Space) instead. The app now tells you this instead of failing silently.
  • Wayland auto-paste requires ydotool (with /dev/uinput access). Without it, text is copied to the clipboard and you paste manually.
  • Bare-modifier hold mode triggers on any press of that key, including inside combos like Ctrl+C. That's inherent to binding a lone modifier; use a combo binding if it bothers you.
  • macOS needs the Accessibility permission for both paste simulation and key-state polling. macOS and Windows builds are untested this release — only the Linux X11 path has been verified end-to-end.
  • Local transcription defaults to English (base.en). Download a multilingual model (tiny/base/small) in onboarding or settings for other languages.
  • LLM cleanup adds latency (sub-second with a warm local Ollama; first request after idle is slower). If the LLM fails or is disabled, the raw transcript is pasted instead — dictation is never lost.
  • Dictation history is capped at the last 100 entries (encrypted at rest, key in the OS keychain).