Highlights
This patch fixes the backend endpoint baked into the build. v0.37.0 artifacts pointed at an unreachable host, so anything account-related — including the new managed Soniox provider — could not connect. Use this release instead. Everything v0.37.0 introduced is listed below.
KizunaAI with Soniox is now available
- A "KizunaAI" provider, powered by Soniox, is now available with no API key. Sign in with your Kizuna AI account and start translating — sessions authenticate automatically and draw on your session balance. (#415, #419)
- 60 languages and 3,600 translation pairs, spoken in your own cloned voice. Record a voice once and managed sessions use it as "My voice".
- Existing KizunaAI selections migrate automatically to managed Soniox instead of silently falling back to bring-your-own-key OpenAI.
Echo is now measured, not guessed
- Sokuji detects echo from the audio itself. The old warning guessed from device names and could not see a real echo; the microphone and participant streams are now correlated against what Sokuji actually played. (#420)
- A floating notice names the specific loop and its fix — translated speech coming back from your speakers, meeting audio reaching the microphone, a participant's device echoing you, the participant source capturing Sokuji's own audio, or an input device that records this computer's playback. It clears itself once the loop is gone. (#420)
- Sokuji no longer lists itself as a participant source, and loopback-style inputs (Stereo Mix, "Monitor of …", VoiceMeeter) now warn when picked and are never selected automatically. (#420)
Provider list
- The list leads with the curated eight: OpenAI, KizunaAI (Soniox), OpenAI Translate, Local Inference, Volcengine AST2, Gemini, Soniox, Palabra AI. (#419)
Full Changelog: v0.36.3...v0.37.2