Improvements
- Kizuna AI voice cloning checks a building voice less often (twice at 1.5 s, then every 3 s), staying within Soniox's voices API rate limit; a build still completes as quickly as before (#499).
- One custom voice per region. Your account now keeps a separate custom voice in each region, so switching regions no longer discards the voice you built in the other one.
Bug fixes
- Kizuna AI voice cloning works in the EU and Japan regions again. Building a custom voice there no longer ends in "Processing failed".
- macOS system audio: the "no permission" warning no longer appears when the only silent output stream is Sokuji's own; the check now excludes Sokuji's own processes and only warns once real tap audio has been seen (#498, fixes #492).
- Audio-capture messages are translated in all 29 supported languages instead of falling back to English (#498).
Note
- Existing custom voices were reset on 2026-09-05. To keep speaking in your own voice, record your reference clip again under Settings → Voice; until then sessions use a built-in voice.
Full Changelog: v0.40.1...v0.40.2