Highlights
OpenAI Realtime: migrated to the GA API — WebRTC works again, voice selection fixed
- Your selected voice is now actually used. Previously the voice picked in settings (Ash, Cedar, …) never reached OpenAI and every session used the default voice. Voice selection now applies consistently on both WebSocket and WebRTC transports. (#373)
- WebRTC connections are restored. OpenAI removed its legacy beta Realtime API on May 12, 2026, which silently broke the WebRTC transport — sessions were quietly falling back to WebSocket. Sokuji now speaks the GA protocol end to end. (#373)
- Repeat sessions stay on WebRTC. Starting a second session no longer fails with an authentication error and downgrades to WebSocket. (#377)
- The conversation panel now displays during WebRTC sessions instead of staying empty. (#377)
- Connection attempts now time out cleanly instead of hanging indefinitely. (#376, #377)
Soniox
- Voice cloning with your own API key — clone a voice and use it for translated speech. (#372)
- 28-voice catalog and session background context for better translation continuity. (#371)
- "Both" shared session: speaker sides are now attributed using diarization, so each side of the conversation lands in the right column. (#370)
- Custom vocabulary, endpoint tuning, and TTS speed settings. (#368)
- The transcription stream now auto-resumes after transient server errors (503) instead of dropping the session. (#374)
Palabra
- Both credential types are now supported: the classic App console client ID + secret pair and the new Platform console
plbr_API keys, selectable in settings. Existing credentials are migrated automatically. (#369)
Full changelog: v0.34.6...v0.35.0