OpenAI has developed GPT-Live, a third-generation voice system that enables more responsive and natural-sounding AI conversations. Previous voice AI systems relied on turn detectors, which introduced delays and could interrupt users. GPT-Live eliminates this separate detector by using a full-duplex voice model that can listen and speak simultaneously.

This new architecture streams audio directly into the voice model and outbound speech back to the user, optimizing for low latency. When complex reasoning or tool use is required, GPT-Live can consult frontier models like GPT-5.5 without pausing the conversation. The system was built over six months, focusing on reworking model inference, context management, and media transport to ensure smooth, end-to-end speech flow.

The architecture separates the core voice path from application logic, allowing for easy customization without impacting responsiveness. This foundation supports new capabilities in ChatGPT Voice, such as computer control and agent coordination in the desktop app. The system leverages WebRTC for low-latency media transport and a stateful inference system for continuous conversation, minimizing audible pauses and artifacts.