Continuous Voice Interaction with GPT Live

OpenAI's GPT-Live is a third-generation, full-duplex voice AI system that eliminates the turn detector from the audio path, allowing the model to listen and speak simultaneously and reducing conversational latency to sub-second response times. Built over six months, the architecture separates the media fast path from application and business logic, enabling tool calls and frontier model delegation (including GPT-5.5) to run asynchronously without stalling audio delivery. The engineering team replaced a Python asyncio implementation with Go, adopted WebRTC for transport, and built a stateful inference system with seamless model-instance handoffs and dynamic context compaction—all keeping the live audio path uninterrupted. The system now powers ChatGPT Voice capabilities including computer control and agent coordination in the desktop app, with a clean customization boundary that lets applications modify tools and policies without affecting media responsiveness.