OpenAI Announces GPT-Live with Full-Duplex Voice Architecture
OpenAI has introduced GPT-Live, a rebuilt voice stack from client to model that enables full-duplex communication, allowing the AI to listen and speak simultaneously. This new architecture keeps audio flowing continuously even when the model is performing deeper reasoning or using tools. This update shifts AI voice interaction from a rigid, turn-based model to a fluid conversation that closely mimics real human dialogue. It significantly reduces latency and allows users to interrupt the AI or receive real-time backchannel feedback like "mhmm" or "yeah". To achieve this, OpenAI routes audio through a dedicated fast path while reasoning and tool use occur asynchronously. Additionally, they reduced the voice-session startup process from six network round trips to just one, utilizing the new GPT-Live-1 and GPT-Live-1 mini models.
## BACKGROUND
Traditional voice assistants operate on a turn-based model where the user speaks, the system processes the audio, and then responds. Full-duplex communication allows simultaneous transmission of data in both directions, enabling real-time interruptions and more natural conversational dynamics.