✨ What's new
Set up subtitle mode before the meeting starts — The subtitle window no longer waits for a running session. Open it whenever you like to position it, size it, and preview your font and colors while people are still joining, then start translating from a button inside the window itself. If something isn't ready — no microphone, a missing API key, models still downloading — the window tells you exactly what is wrong, and one click takes you to the setting that fixes it. (Desktop app; #324)
🛠 Fixes & polish
- A broken default microphone no longer hides every other device. When the system default input was a stale or phantom device — common after replugging a USB mic — Sokuji froze for about 20 seconds at launch and then reported "no audio devices" until a reboot. It now lists your devices right away and keeps working with the one you picked.
- On-device Voxtral no longer clips the end of a sentence. The last word or syllable of each utterance was being cut off, which was especially obvious with Japanese sentence endings. Tails are now decoded in full, and each pause is turned around faster than before.
- New Silence Threshold control for on-device speech detection. Lowering the VAD Threshold to pick up quiet speech could stop segments from ever ending — they ran to the 20-second cap and were cut mid-sentence. The silence threshold now follows the speech threshold automatically, and you can set it yourself under VAD Settings.
- Soniox now tells you when spoken translation stops. A failure in the speech output used to be silent: captions kept scrolling with no audio and no explanation. You now get a clear message that says transcription and text translation are still running, and ordinary reconnections after a pause no longer look like errors.
- Updated dependencies to clear a batch of security advisories.
Full Changelog: v0.34.2...v0.34.5