How OpenAI achieved low-latency voice AI at scale: relay + transceiver architecture

Friends, I want to share news from the OpenAI ecosystem: engineers described how they reduced voice-AI latency at global scale.

What it covers:
- Problems: port exposure and ICE/DTLS state "sticking" when scaling on Kubernetes.
- Solution: split into a lightweight relay (UDP forwarding) and a stateful transceiver, route by ICE ufrag, global ingress points.
- Gain: smaller public UDP surface, geo ingress for a short first hop, and retention of standard WebRTC for clients.

Why it matters: enables live, low-latency voice interactions while simplifying security and scaling.

What do you think of this architecture for your realtime services?

#OpenAI #WebRTC #VoiceAI #Infrastructure