OpenAI has released details on GPT-Live, a new voice AI system built for real-time conversation. The company says the project took six months to build.
GPT-Live is OpenAI's third generation voice system. It powers ChatGPT Voice, including a new feature that lets people control their computer through voice commands.
The system is described as full-duplex. This means it can listen and speak at the same time, much like two people talking naturally.
Older voice AI systems used a turn-based design. They relied on small programs called turn detectors to guess when a person finished speaking.
Those detectors often made mistakes. Guessing too soon could cut off the user, while guessing too late made the response feel slow.
How the System Works
GPT-Live removes the turn detector entirely. The voice model itself decides when to speak, based on continuous audio flowing in and out.
When the system needs deeper reasoning or tool use, it can call on other models, such as GPT-5.5. This happens in the background so the conversation does not pause.
Engineers separated the audio path from the application logic. This means a slow tool call or backend service cannot stall the flow of speech.
The team rewrote the media system in the Go programming language. It had previously used Python, and the switch improved how smoothly audio frames were delivered.
GPT-Live uses WebRTC to move audio between the user and the model. WebRTC is designed to keep working even when there is packet loss or a change in network connection.
Faster Connections and Testing
OpenAI also built a new protocol called WARP, short for WebRTC Abridged Roundtrip Protocol. It cuts the number of network round trips needed to start a session from six down to one.
The company worked with other developers in the WebRTC community on WARP. Support for it has already been added to two widely used WebRTC software libraries, libwebrtc and Pion.
Before releasing GPT-Live to users, OpenAI ran a silent test. A small share of real ChatGPT Voice sessions were routed through the new system without changing what users actually heard.
The test showed that speed depends on more than just processing power. Other parts of the system, such as network paths and stream handlers, had to scale along with it.
Long voice sessions created their own challenges. The company said reconnects and extended calls exposed issues with memory use and saving conversation states.
OpenAI plans to expand GPT-Live further. The company said it will support an upcoming GPT-Live API and expand voice features across more devices and apps.
