Building Voice Applications with GPT Live 1: Full-Duplex Speech, Delegation, and Turnless Architecture

A live DevDay session introducing GPT Live 1, a full-duplex voice model that streams audio in and out simultaneously and keeps conversing while a backend agent performs tasks. The presenters run three demonstrations to explain the core concepts needed before building with the model, and they operate the entire session by voice.

■ Model Personality and Prompting
- Full-duplex listening and speaking, with turn-taking handled by the model instead of a separate voice activity detector
- Promptable tone, accent, back channels, laughter, patience when interrupted, and whether the model speaks at all
- Running a second session as a conversation partner or as a way to build evaluation scenarios

■ Delegation and Backend Coordination
- Two delegation options in the API: client delegation and responses delegation
- Three channels for backend updates: silent context, speakable content, and instructions
- Deciding update frequency and detail level as a product decision

■ Turnless Architecture
- Continuing to listen and respond while loud background audio plays
- Using voice to surface niche settings and collect richer context from users

Developers building voice applications will find concrete architecture choices, prompting strategies, and migration considerations that let them keep an existing agent harness and simply add a voice interface. The session ends with pointers to a starter app and developer plugin so viewers can begin experimenting right away.