Skip to content

v0.35.2

Choose a tag to compare

@github-actions github-actions released this 09 Aug 05:45
· 2 commits to main since this release

Highlights

Capture one application instead of the whole system

  • Participant audio can now come from a single application. Pick the meeting app — or whatever is actually playing the speech you want translated — and Sokuji captures only that, instead of everything your machine happens to be playing. Notifications, music, and your other tabs stay out of the translation. Whole-system capture remains the default, so nothing changes unless you choose an application. (#380)
  • Works on all three desktop platforms, each through its native path: WASAPI process loopback on Windows, a Core Audio process tap on macOS, and PipeWire on Linux.
  • Switch sources mid-session. Changing the captured application works the same way the microphone hot-swap does — no need to stop and restart the session. Your choice is remembered across restarts, matched to the app itself rather than to a process that won't exist next time.
  • If capturing a specific application fails, Sokuji falls back to whole-system capture and now tells you it did — that fallback widens what reaches your translation provider, so it should never happen silently.

Soniox

  • Outages now end with a sentence that makes sense. A session that dropped for a reason you didn't cause used to end on the server's raw English error — [Soniox 503] service unavailable — or, for a plain network drop, on nothing at all. These now end with a localized, actionable message telling you what happened and what to do next. (#383)
  • The raw server text is still kept where it is useful, in the Logs panel timeline, for when you need to diagnose something rather than just get on with your call.
  • Errors that genuinely need the server's own wording — authentication failures, rate limits, bad requests — still show it verbatim, because there the server's message is the actionable part.
  • Fixed: stopping a session could run its teardown twice, and a session ending late could tear down whichever session had started after it.

Linux

  • Fixed: on Wayland the app started but no window ever appeared. On a native Wayland session (GNOME on Ubuntu, among others) Sokuji would launch, run in the background, and never show anything — no window, no error, nothing in the window switcher. The cause was a GPU setting that Chromium cannot honour on Wayland; it left the renderer unable to produce a single frame, so the window was never actually created. (#390, reported in #389)
  • The .deb now launches through XWayland, which keeps GPU acceleration available for local inference on Wayland desktops. xwayland is listed as a recommended package so it is installed by default.
  • Local inference tells you when it is running without a GPU. If Sokuji falls back to CPU rendering, the Models section now says so once, and on Linux it shows the flag that restores hardware acceleration. Models still run in this state — just noticeably slower.
  • Fixed: virtual audio device setup failed when pactl was missing. The .deb now recommends pulseaudio-utils, which provides it.

If you run the AppImage on Wayland, or start the .deb from a terminal rather than your app menu, local inference will use CPU rendering. Launching with --ozone-platform=x11 restores GPU acceleration.

Full changelog: v0.35.1...v0.35.2