Qwen3 TTS has input streaming mode: you can stream LLM output into the speech model. So don’t need to wait for a full sentence. We also implement this websocket variant, and it also runs at sub 50 ms.

LLM TTFT is still a big issue, and we might tackle that problem as well.