Summary
OpenAI’s Aarti answers developer questions about GPT-Live, the company’s new voice model series, in this community Q&A. The video begins with a short history of voice AI: early cascaded systems chained speech-to-text, a language model and text-to-speech, adding latency at every step, while the GPT Realtime models merged everything into a single speech-to-speech model that remained largely turn-based.
GPT-Live introduces two architectural changes. The first is full duplex, meaning the model listens and speaks at the same time, so it can tell a simple “mm-hmm” acknowledgement from a real interruption and handle background noise or side conversations more gracefully. The second is delegation: the live model manages the conversation and hands complex work to a backend model or system, which lets developers swap in newer backend models or add voice to existing text applications.
The session also covers real-world deployments, including Yelp for reservations and food orders, Fin for customer support, Speak for language lessons, and Cognition for voice in Devin. Aarti demonstrates new French voices (Brise and Sillage), explains custom voice creation through prompt-to-voice and recordings with consent, and offers troubleshooting advice on delegation, including keeping the GPT-Live prompt short, using the SpawnThinking token with GPT-Live 1, and forcing delegation through Decisions triggers.
📺 Source: OpenAI · Published November 10, 2026
🏷️ Format: Interview
