2 hours ago
1
OpenAI is updating its AI voice technology to match its text-based artificial intelligence. Two new models, GPT-Live-1 and a mini version, are rolling out now to all ChatGPT users.
The new model duo relies on OpenAI's latest frontier models. You can select from three different levels of intelligence, depending on how in-depth you want the responses to be. But the biggest update should be that the voice sounds more human and is more conversational, Atty Eleti, product lead for ChatGPT voice, told reporters.
One of the biggest changes under the hood is a new process called continuous interaction. This framework allows the AI to simultaneous receive information and produce outputs -- so in the case of voice, it can "listen" and "speak" at the same time. This is a change from previous voice modes that could only respond after you finished speaking.
It's helpful if you want to do live, simultaneous translation: You can speak in English and have ChatGPT translate what you're saying into Spanish, Hindi or another language with only a short delay.
Part of the new architecture can also delegate or offload tasks to OpenAI's frontier models. For example, it can answer a question you asked while researching another one.
"When GPT-Live has to think hard for a question, it can delegate its reasoning and complex task to GPT-5.5, which can do things in parallel, and this GPT-Live can still remain in conversation with the user," said Kundan Kumar, research lead for the GPT-Live model. When it's done, it "seamlessly weaves" the answer into its response.
"This is exactly how humans interact with each other," Eleti added. "We keep the conversation going while we think in the background."
It also uses OpenAI's newer AI design tech to create some visual answers, when appropriate, "because sometimes the best answer is displayed, not spoken out loud," Eleti said. Think weather reports, sports scores and other kinds of graphics.
This is an example of a graphic ChatGPT would make if you asked it to give you a weather report while in voice mode.
OpenAIThese new voice models are now available to all ChatGPT users, including free users and paying subscribers, on mobile apps and on the website. Don't worry if you don't see it right away; OpenAI says it may take a day or two to ensure everyone has access. This is separate from Thursday's expected GPT-5.6 series drop.
You'll notice that one way the AI voice tries to mimic human speech is with filler words -- those "ums," "ers" and "likes" that we use while thinking through our thoughts aloud. Those are new to this update. You can also interrupt it with fewer lags. You can have it sit quietly and listen, only responding when called upon with its name as a wake word. Unlike Siri and Alexa, though, ChatGPT will actively listen until it's manually disabled.
OpenAI automatically opts you out of AI training with voice mode. Audio clips are stored for 30 days -- so the AI has the context from prior conversations -- and can be deleted.
Having an AI that feels more human comes with risk, though. We've seen numerous cases (and lawsuits) of the harms of anthropomorphizing AI can wreak, particularly for people struggling with their mental health. OpenAI says the new models have expanded safeguards and performed better than previous models on "key safety areas" including self-harm, psychosis, violence and sexual content.
