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

OpenAI has introduced a version of its latest real-time, full-duplex voice model, GPT-Live-1, as an API for software developers. This allows for more responsive voice-enabled apps and conversational workflows. The model can listen and generate speech simultaneously, leading to fluid communication and reducing the frustration of speakers and listeners taking turns, as experienced with half-duplex handheld radios.

OpenAI has prioritized customization options for developers implementing GPT-Live-1, enabling them to tailor voice interaction experiences and associated workflows for business applications. Early evaluations have shown that GPT-Live-1 significantly reduces interruptions during conversations, with Speak reporting an 80 percent decrease in interruptions compared to previous turn-based systems.

GPT-Live-1 is designed to handle spoken conversation while a backend model, such as GPT-6 Astra or a more cost-effective option, manages information lookup, tool usage, and task management. The model has shown impressive performance, scoring 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 claims that GPT-Live-1 outperforms GPT-Realtime-2.1 in Full Duplex Bench performance by 30 percent, while also improving turn-taking latency and interactive behavior. Yelp, which has been using GPT-Live-1 for its AI restaurant reservation system, reports that it makes every call more conversational and responsive, enhancing guest experiences and freeing up staff to focus on serving customers.

Pricing for GPT-Live-1 is set at $0.05 per minute of usage, in addition to any backend models being utilized. Custom voice options can be arranged with OpenAI's sales team. The voice model is also available through OpenAI Presence, the company's enterprise agent platform.