Simplifying Voice Architecture

Traditional voice agents rely on cascaded architectures—stitching together separate Speech-to-Text (STT), Large Language Model (LLM), and Text-to-Speech (TTS) components. This approach introduces significant latency and makes it difficult to manage natural conversational flow, such as handling interruptions or pauses.

GPT-Live-1 replaces this with a single, full-duplex model that processes incoming and outgoing audio simultaneously. This architecture allows the agent to handle interruptions and acknowledgments in real-time. By offloading the voice layer to GPT-Live-1, developers can delegate deeper reasoning and tool-calling tasks to specialized backend models (like GPT-6 Astra or Terra), enabling the conversation to continue fluidly while the backend processes complex requests.

Performance and Control

GPT-Live-1 offers significant improvements in interactive behavior, including a 30 percentage point improvement in Full Duplex Bench performance compared to previous iterations. Key operational advantages include:

  • Interruption Handling: The model reasons over audio holistically, reducing interruptions by nearly 80% in early testing (e.g., in language tutoring scenarios).
  • Context Management: Improved handling of background noise and silence, preventing the model from narrating every environmental sound.
  • Developer Customization: Developers can steer tone, pace, and style via system prompts, and the model supports alphanumeric understanding and keyword biasing.
  • Deployment Flexibility: The model supports telephony and extended sessions, making it suitable for high-volume tasks like customer support, reservations, and AI-assisted engineering workflows.

Integration and Availability

The API is available at $0.05 per minute for the voice layer. Developers can integrate it by passing conversation context to a backend model and returning the response to the GPT-Live-1 session. The platform also supports OpenAI Presence for enterprise-grade deployments that require integration with proprietary company systems and human escalation paths.