OpenAI introduced GPT-Live, a new voice AI system designed to enable continuous, real-time conversations by eliminating the need for turn-based processing. The system allows for full-duplex interaction, meaning it can listen and speak simultaneously, making conversations feel more natural and immediate. This marks a significant shift from previous voice AI systems that relied on turn detectors to manage speech turns, often leading to delays or awkward pauses. GPT-Live integrates with frontier models like GPT-5.5 to handle complex tasks without disrupting the flow of dialogue, enhancing both responsiveness and intelligence in voice interactions. Source: openai
GPT-Live was developed over six months with a focus on low-latency architecture and seamless media flow. The system streams audio directly into the voice model and sends speech back to the user while handling deeper reasoning and tool use asynchronously. This approach ensures that the conversation remains uninterrupted, even when invoking advanced models or managing context. The architecture also separates media flow from application logic, allowing for customization without affecting the responsiveness of the voice system. By using WebRTC for transport, the system maintains low-latency performance and can recover from packet loss or clock drift, ensuring smooth audio delivery. Source: openai
The development of GPT-Live was driven by the limitations of earlier turn-based systems, which struggled to keep up with the fluidity of human speech. Previous architectures processed speech-to-text, then the LLM, and then text-to-speech in a series, adding latency and missing contextual cues. GPT-Live addresses this by processing audio directly, preserving speech details and enabling faster responses. The system also introduces stateful inference, allowing for continuous context management and dynamic compaction of conversation history to maintain performance without interrupting the user experience. Source: openai
Source: openai
