SGLang-Omni 的流式音频

Problem

When a model speaks out loud, the audio has to keep up with the listener. A gap of a few hundred milliseconds is not a slow page, it is a stutter in someone’s voice. Serving frameworks are tuned for tokens per second; speech needs a chunk of sound ready before the previous one finishes playing.

What I worked on

• Seeing the gap. The request timeline showed when a codec frame reached the vocoder and when audio reached the client, but nothing in between. I added decode-step events, so a late chunk can be explained instead of guessed at.
• Letting urgent work cut the line. With those events it became clear that follow-up chunks were waiting behind work that nobody was listening to yet. Urgent chunks now launch first and commit when they finish.
• Not letting one slow listener win. Each stream held an unbounded queue of finished audio. One client that read slowly could keep growing memory after its generation was done. That queue now has a ceiling.

Why it’s interesting

All three are the same lesson in different clothes: in a streaming system, finishing fast is not the goal. Arriving on time is. It is the same question my research asks of translation, one layer further down.