OpenAI is rolling out two new voice models—GPT-Live-1 and a smaller variant—to all ChatGPT users, bringing the service’s spoken interactions closer to the fluidity of its text-based AI. The update, available on mobile and web, introduces a feature called continuous interaction, which lets the AI process incoming audio while generating speech in real time. Product lead Atty Eleti told reporters the result is a more natural, human-like conversation, with the system able to listen and respond without waiting for a pause.

Under the hood, the new architecture can delegate complex reasoning tasks to OpenAI’s frontier models, like GPT-5.5, while keeping the main voice channel open for ongoing dialogue. Research lead Kundan Kumar explained that when the AI needs to think through a tough question, it can hand off the heavy lifting to a separate model and later weave the answer back into the conversation seamlessly. This allows for live translation—speak in English and hear Spanish output almost instantly—and lets users interrupt or ask follow-ups without lag. The voice now includes filler words like “um” and “like” to mimic human speech, and a wake-word feature lets ChatGPT sit silently until called.

OpenAI has added expanded safety guardrails, with the new models scoring better than predecessors on preventing harmful responses related to self-harm, violence, and sexual content. Audio clips are stored for 30 days to maintain conversation context but are not used for AI training by default. The update is separate from the anticipated GPT-5.6 release later this week, and OpenAI warns it may take a day or two for all users to see the new voice options.

Source: CNET