gpt-realtime

GPT-Realtime is OpenAI’s unified, speech-to-speech model: it enables ultra-low-latency, human-quality voice agents for real-time conversational AI.

This general-availability model (gpt-realtime) delivers seamless, 'speech in, speech out' interactions, eliminating the latency of traditional transcription-LLM pipelines. It uses a single, unified architecture: this allows for natural conversational flow, including handling interruptions and emotional nuance. Developers connect via low-latency protocols (WebRTC, WebSocket, SIP) to build high-performance voice applications. Key use cases include customer support, where agents require immediate response, and educational tutoring, leveraging the model’s 32,000-token context window for stateful, continuous dialogue.

https://platform.openai.com/docs/api-reference/realtime