Highlights
Gemini Live Translate
gemini-3.5-live-translate-previewnow works properly when you pick it from the Gemini model list. It is Google's dedicated interpreter model — built for continuous low-latency interpreting rather than turn-by-turn replies, and it carries the speaker's own voice through the translation instead of substituting a synthetic one. Choose it and set your languages as usual. (#398)- The target language is now set through the API's own translation field instead of being inferred from the system prompt. That inference held for a short, direct prompt and broke on a longer one: the translation could come out in an unrelated language and the conversation log and subtitles went blank at the same time, with nothing on screen to explain either. Advanced mode was the most exposed, since it hands you the entire prompt and never substitutes your chosen languages into it. (#398)
- Your system prompt still applies, and remains the way to give the model terminology or style guidance — naming proper nouns in it measurably improves how they come out.
- Controls this model ignores are now hidden for it: voice, temperature, and max tokens. It reproduces the speaker's voice by design, so a voice picker was only ever offering a choice that did not exist.
- "Auto Detect" as your source language is now refused, with an explanation, when you are also translating other participants. That direction has to be reversed to translate into your language, and "auto" is not a language to reverse into. Pick a concrete source language and the session starts. (#398)
macOS audio capture
- Fixed captured system audio silently going dead when the target application recycles its audio process — something Chrome does routinely. Nothing looked wrong: the session stayed healthy and kept delivering perfectly clocked buffers of pure silence. Sokuji now tracks the application's audio processes for the whole capture instead of resolving them once at the start. (#395, closes #393)
- Capture no longer refuses an application that has not started playing audio yet. It attaches straight away and picks the audio up when it appears.
Under the hood
- Upgraded
@google/genaito 2.x, which the Live Translate support above requires. (#397)
Full Changelog: v0.35.5...v0.36.0