OpenAI today began rolling out Advanced Voice Mode to a small group of ChatGPT Plus subscribers, delivering the real-time, emotion-aware voice interface first demonstrated with GPT-4o in May. The alpha arrives roughly a month after OpenAI said it needed additional safety work, and it lacks the video and screen-sharing capabilities shown in the original demo.

The new mode is a departure from ChatGPT’s previous voice pipeline, which used three separate models to process speech. GPT-4o is multimodal by design, handling transcription, reasoning, and voice generation in a single step — dramatically reducing latency. OpenAI also says GPT-4o can sense emotional intonations in a user’s voice, including sadness, excitement or singing.

For now, only a handful of Plus users receive an in-app notification and email with instructions. OpenAI plans a broader rollout to all Plus subscribers by fall 2024. The alpha is limited to four preset voices — Juniper, Breeze, Cove, and Ember — created with paid actors. The Sky voice, which drew controversy after Scarlett Johansson said it resembled her voice without consent, is no longer available. The company says it has deployed filters to block impersonation of public figures and to prevent generation of copyrighted audio.