Highlights
Soniox can run in the EU or Japan
- Added a Region selector for Soniox (EU, Japan, or global). Each region is a separate Soniox deployment with its own API key, so bring-your-own-key now stores a key and a voice per region. (#413)
- Your audio is processed and stored in the region you pick, from the next session onward. Changing region re-validates your credentials instead of reusing the previous result. (#413)
- Managed sessions use the region the backend confirms, so a settings change mid-connect can never pair one region's key with another region's host. (#413)
A session that is still starting can be cancelled
- Start is no longer a one-way door — both footers can cancel while voice preparation or model loading is still running. (#408)
- A cancelled start leaves nothing behind: no session activates, microphone capture stops, and no initialization error is reported. (#408)
- The advanced footer now shows the
Loading model…step that the simple footer always had. (#408)
Audio and transport
- The microphone waveform now moves during WebRTC sessions. These sessions never start the shared recorder the waveform read from, so the input strip previously sat flat while audio flowed normally. (#410)
- Each provider's transport setting now governs only that provider. Choosing WebRTC under OpenAI used to decide how every other WebRTC-capable provider started its session. (#410)
Participant translation direction
- Invalid reversed language pairs are now rejected before connecting. Zoom AI, Volcengine ST, Palabra AI and OpenAI Translate reverse the direction for the participant leg without checking the result against their own catalog, so an unsupported pair only surfaced as an API error after the session connected. (#411)
Under the hood
- Session start now dispatches through per-provider descriptors instead of branching on provider identity inside the main panel, pinned by golden-equivalence tests against the previous implementation. (#408)
Full Changelog: v0.36.2...v0.36.3