OpenAI arms devs with AI conversation tool that can talk and listen at the same time

OpenAI has introduced GPT-Live-1, an API version of its latest real-time full-duplex voice model, allowing developers to create more responsive voice-enabled applications and conversational workflows. This full-duplex model can listen and generate speech simultaneously, facilitating smoother communication compared to previous half-duplex systems. Initially available within the ChatGPT interface, GPT-Live-1 is now accessible through API calls.

The model emphasizes customization for developers, aiming to enhance business app experiences with voice interaction and associated workflows. In initial evaluations, GPT-Live-1 demonstrated significant improvements over its predecessors, such as Speak, which reported an 80 percent reduction in interruptions when compared to previous turn-based systems.

GPT-Live-1 is designed to work alongside backend models, like GPT-6 Astra or more cost-effective alternatives for routine tasks, handling information retrieval, tool usage, and task management. Its performance surpasses that of OpenAI's Realtime API, with scores of 86.2 percent in the Tau3 voice-agent intelligence benchmark, compared to 45.7 percent for GPT-Realtime-2.1 and 42.4 percent for GPT-Realtime-2.

OpenAI cites its evaluations, claiming GPT-Live-1 outperforms GPT-Realtime-2.1 by 30 percent in Full Duplex Bench performance, while also excelling in turn-taking latency and interactive behavior.

Yelp has integrated GPT-Live-1 into its Yelp Host service, expressing satisfaction with the technology's impact on restaurant reservation systems. The voice AI enhances conversational and responsive interactions, allowing restaurants to focus on serving guests rather than handling phone calls. OpenAI charges $0.05 per minute of GPT-Live-1 usage, with the potential to negotiate custom voices through its sales team. The voice model is further accessible through OpenAI Presence, the enterprise agent platform.