SourceHow we built a realtime system for responsive voice AI in six months
Users experience sub-second responsiveness in ChatGPT Voice, enabling natural interruptions during conversation. This technology also powers new capabilities like controlling computers and coordinating agents in the desktop app.
- Eliminates turn detector with a full-duplex model that listens and speaks simultaneously
- Delegates complex reasoning and tool execution asynchronously to GPT-5.5
- Rewrote media frontend in Go, bringing p95 latency down to previous p50 levels
- Developed WARP protocol with IETF to cut WebRTC startup handshake from six round trips to one