OpenAI has recently officially announced the introduction of the new model GPT-Live-1 to its API, providing core support for developers to create full-duplex real-time voice applications and business processes. The launch of this model marks the evolution of AI voice interaction from traditional "segmented processing" to true natural conversation.
In terms of core technical architecture, GPT-Live-1 achieves deep integration of speech understanding and speech output within the same model. It breaks through the multiple connection barriers in traditional models that involve "speech-to-text, large language model reasoning, and text-to-speech," thereby significantly reducing system latency. This model fully supports full-duplex conversation, allowing the system to listen to the user's voice while outputting speech, and accurately handle complex scenarios such as interruptions, pauses, and background noise during the conversation.
In addition to a smooth interactive experience, GPT-Live-1 also demonstrates high flexibility in function expansion and scenario implementation. Developers can easily adjust the model's tone, speed, and conversation style through system prompts, and configure backend models, tools, and intelligent agent frameworks. At the same time, the model supports native speech recognition and transcription, with capabilities including alphanumeric understanding, keyword bias, and round detection. It can connect with tools like Codex, and through OpenAI Presence, it can develop voice work that queries enterprise systems, performs approved operations, and transfers to human agents when necessary. In practical application scenarios, it can seamlessly integrate into phone conversations, handling full-duplex voice intelligent agent tasks such as restaurant reservations and customer service.
The effectiveness in practical applications is also significant. During early evaluations, the language learning platform Speak saw an almost 80% reduction in mis-interruptions during learners' thinking pauses after integrating GPT-Live-1, compared to previous round-based systems. Medical service platforms have also streamlined their architectures and significantly reduced code volume using this model. According to OpenAI's evaluation, GPT-Live-1 outperformed GPT-Realtime-2.1 by 30 percentage points in the Full Duplex Bench test, and ranked first in the Tau3 end-to-end voice intelligent agent task test when combined with GPT-6Astra at a moderate reasoning intensity.
