- Chinese is the least reliable target. With identical audio feeding four to six sessions, the zh-Hant session is the one most often born mute, thinned, or refused, while ko, vi, es and en on the same input stay up.
- Mandarin spoken into a Chinese target silences the session. With echoTargetLanguage: true, when the speaker switches into Mandarin the zh-Hant session stops producing audio and transcription for tens of seconds while other targets keep translating.
- Quota refusals arrive only as socket closes. With six concurrent sessions on one project, new connections and goAway resumes were closed with 1011 “You exceeded your current quota” dozens of times within minutes. No advance signal, and no published concurrent-session or per-minute limit for Live translation, so an event cannot be sized.
- Sessions are born mute. A new session accepts audio and produces input transcription, yet returns only silent frames and no output transcription. Nothing marks it as broken; the client must detect it by listening and reconnect.
- Output stalls, then arrives in a burst. Input transcription keeps streaming while translated audio slows to a trickle for one to three minutes, then a hundred-plus output chunks land in a few seconds. Our send cadence is steady. No busy or backpressure event is sent.
- Quality differs per session on identical audio. The same input yields fluent, choppy, or near-silent output depending only on which session was drawn, worst on Chinese. Reconnecting re-rolls it, so quality cannot be tested or promised.
- 1011 internal error closes under load, with no reason. Session deaths rise with the number of concurrent sessions on a project. Nothing in the close says whether it was capacity, a server fault, or the client.
- Audio and transcription desynchronize. Sessions emit output transcription with no audio, or audio with transcription gaps, so captions and speech disagree about whether translation is happening.
- Translated Chinese lags and chunks poorly. Traditional Chinese output arrives in long delayed segments rather than phrase by phrase, so captions jump and the audio trails the speaker more than other targets do.
- No signal when a session degrades. Between “healthy” and “closed” there is no event for starvation, throttling, or a mute generator. Every failure above has to be inferred from silence.
- Resumption handles are not always honoured. Some resumes after goAway come back cold, with no continuity, and nothing in the response says the handle was rejected.
I know this is still in preview, but I think this is too unstable at the moment. Any time frame approximation of next generation or stable release? I really hope our team won’t forced to abandoned this project.