OpenAI has shipped a new generation of voice models for ChatGPT. The demo features four engineers, Kundan Kumar, Yuchen Zhang, Ehsan Asdar, and Rithesh Kumar, walking through what the system can actually do in real time, not in a polished promo reel.
The engineers are the story here. Researcher-led demos carry technical credibility that marketing videos do not. Watching the people who built the system stress-test it live reveals failure modes, latency characteristics, and capability ceilings that a press release will never touch.
Read the full transcript and watch the demo for the specific model architecture decisions, the latency numbers, and the edge cases they choose to demonstrate. Those choices are never accidental.
[WATCH ON YOUTUBE →]