How we built a realtime system for responsive voice AI in six months

OpenAI · 2 months ago · 55 ● 2 sources

A team built GPT-Live, a voice AI system that processes speech continuously and supports full-duplex conversation without waiting for turn detection. The system achieved low-latency performance through streaming audio and asynchronous processing in a six-month development cycle. Users can now have uninterrupted back-and-forth dialogue with the AI without the delays of traditional turn-based systems.