OpenAI upgrades ChatGPT Voice with GPT-Live-1 model
OpenAI says the new voice model can listen while speaking, cut down on interruptions and translate speech in real time.
By Maya Lindqvist · Senior Technology Correspondent
3 min read
OpenAI is rolling out a new model for ChatGPT Voice that it says will make spoken conversations feel less stilted and reduce unwanted interruptions. The update matters because voice assistants often struggle with pauses, overlapping speech and the timing cues that shape ordinary conversation.
The company calls the model GPT-Live-1, and OpenAI researcher lead Kundan Kumar described it during a press briefing as the company’s “smartest voice model” so far. OpenAI says the model can keep listening while it talks, allowing ChatGPT to respond more fluidly when a user pauses, continues speaking or cuts in.
OpenAI product lead Atty Eleti said during the briefing that GPT-Live-1 is a “full duplex model.” Eleti said that means the system can process incoming audio and generate outgoing speech at the same time, rather than waiting for one side of the exchange to finish before the other begins.
That is a change from the older ChatGPT voice setup, which OpenAI said used a turn-based approach. According to OpenAI, the previous model could give inaccurate answers and had trouble maintaining the flow of a spoken exchange.
GPT-Live-1 can also send a user’s request to OpenAI’s stronger text models, including GPT-5.5, when the system needs to reason through a question or search the web, OpenAI said. The company says that handoff should help ChatGPT move more quickly from gathering information to speaking back to the user.
OpenAI said the new voice mode can add AI-generated visual material to some conversations. For topics such as weather, stocks and sports, the company said ChatGPT may show information such as forecasts or scores alongside the spoken response.
New voice features
The duplex design enables real-time translation in ChatGPT Voice, according to OpenAI. Instead of waiting for a person to stop talking, the system can translate as the person speaks.
OpenAI said users can also tell ChatGPT Voice to stay quiet until they address it again, a behavior the company says was not available before. While listening, the model may acknowledge the user with short responses such as “mhmm,” “yeah” or “got it,” according to OpenAI.
The company has framed the update as an effort to make ChatGPT feel closer to speaking with another person. The main changes center on timing: fewer interruptions, better handling of pauses and the ability to listen while producing speech.
Safety and availability
OpenAI said GPT-Live-1 includes safeguards meant to prevent harmful answers and, in higher-risk cases, end conversations. The Verge has reported that OpenAI is facing lawsuits alleging ChatGPT contributed to delusions and harmed users’ mental health.
OpenAI said the model is trained to provide “expert-vetted crisis helpline support” in conversations involving self-harm. The company also said it is designed to give “age-appropriate” responses to teenagers.
OpenAI said GPT-Live-1 is rolling out on iOS, Android and the web. The model will run ChatGPT Voice for Go, Plus and Pro subscribers, while free users will get GPT-Live-mini-1, a smaller and more efficient version, as the default.
This story draws on original reporting from The Verge.