GPT Live Q&A: Architecture, Delegation, and Voice Features Explained

A Q&A session covering GPT Live, answering viewer questions on how it compares with the GPT Realtime series, and explaining its key architectural changes. The discussion includes practical guidance on prompting, delegation settings, custom voices, and context management for developers building voice agents.

■ GPT Live Architecture
- Evolution of speech models from the sequential pipeline to GPT Realtime and GPT Live
- Full-duplex listening and speaking, plus delegation to a backend model
■ Delegation & Prompting
- Preventing skipped delegations and keeping the frontend prompt concise
- Using Decisions API triggers to enforce delegation
- Avoiding over-delegation when collecting tool call parameters
- Client-side delegation for precise control over context
■ Use Cases & Voices
- Real-world deployments such as Yelp, Finn, Speak, and Cognition
- New multilingual voices with regional accents, and custom voice generation options
■ Context Management
- Automatic context compression and best practices for long conversations
- Two delegation modes and how they affect context passed to the backend

This is aimed at developers building voice agents or migrating from GPT Realtime to GPT Live. Viewers can gain practical knowledge on configuration, prompting strategies, and context handling to improve their implementations.