Real-time avatar: startAvatarAsync hangs ~29 s then fails "Internal server error. websocket error code: 1011" in bursts, across regions

Hello

Intermittently, users complained a massive lag when the live avatar replies the 1st time.

So, we did various tests

1. Resources: Speech S0 in Central India and Southeast Asia . Client: JS Speech SDK from aka.ms/csspeech/jsbrowserpackageraw, headless Chromium on AWS Lightsail (India). Setup follows Microsoft's samples/js/browser/avatar: relay token from /cognitiveservices/avatar/relay/token/v1, TURN only, eventChannel data channel, voice set in SSML only.
2. Test on 2026-11-12: 13 standard video avatars (lisa, lori, meg, harry, max, in their styles), 8 rounds, both regions, 35 s between starts to respect 2 new connections per minute. 208 starts in total.
3. 18 failures, all between 18:42 and 18:59 UTC (00:12–00:29 IST). Each failure: startAvatarAsync resolves Canceled after about 29,000 ms with Internal server error. websocket error code: 1011. ICE never starts and no SDP answer arrives.

In 14 of the 18, the same avatar failed in both regions within 35 s of each other.

Every other start succeeded in 1.3–1.8 s, including all 190 starts from 18:59 to 20:23 UTC and 8 more on 2026-11-12 at 05:00 UTC.

Real users see the same pattern: the first start hangs about 30 s and fails with 1011, then a retry about 60 s later connects in under 2 s.

QUESTION: what causes these bursts that affect both regions together (backend capacity or scaling?), how can we avoid them, and is there a status or health signal we can check before starting a session?

This kind of a bad start is a HORRIBLE user experience as it happens at the 1st response from the Avatar, intermittently, & quite unacceptable. Other non-Microsoft providers don't have such issues with their AI avatars! Full logs available on request, offline if needed