ChatGPT's New Voice Model Sounds More Human, but There's a Risk

JAKARTA - OpenAI has started to launch two new voice models for ChatGPT, namely GPT-Live-1 and the mini version. This update makes ChatGPT able to hear and speak at the same time, including for direct translation.

CNET quoted Thursday, July 9, saying that the two new models are available to all ChatGPT users, both free users and paid subscribers. Access is provided through a mobile app and website.

OpenAI is updating its ChatGPT voice technology to bring it closer to its text-based AI capabilities. Users can also choose three levels of intelligence, depending on the depth of the answer needed.

Atty Eleti, the leader of the voice product ChatGPT, said the biggest change was in the voice that sounded more human and felt more like a conversation.

One of the key technologies behind this update is called continuous interaction. This system allows AI to receive information and generate answers at the same time.

In voice mode, ChatGPT can "listen" and "talk" at the same time. This is different from the previous voice mode which only responds after the user has finished speaking.

This ability is important for direct translation. Users can speak in English, then ChatGPT translates their speech into Spanish, Hindi, or other languages with a short pause.

This new architecture can also divide tasks into OpenAI's frontier model. The frontier model refers to the most advanced AI model used to handle more complex tasks.

Kundan Kumar, the lead researcher on the GPT-Live model, said GPT-Live can keep up a conversation with a user while handing off complex reasoning to GPT-5.5.

"When GPT-Live has to think hard about a question, this model can delegate its reasoning and complex tasks to GPT-5.5, which can do many things in parallel, and GPT-Live can still continue to talk to users," said Kumar.

Once that process is complete, the answer will be strung back into a conversational response.

"It's exactly how humans interact with each other," said Eleti. "We continue to carry on the conversation while thinking behind the scenes."

This new voice model can also display visual answers in certain situations. According to Eleti, sometimes the best answer is not spoken, but displayed. For example, weather reports, sports scores, or graphs.

OpenAI said the launch was phased. If users have not seen this feature, access can appear in one or two days. This update is separate from the launch of the GPT-5.6 series, which is expected to arrive on Thursday.

There are also changes in speech style. The AI voice can now use filler words such as "um", "er", and "like", like a person who is thinking while talking.

Users can also interrupt the AI's response with a shorter pause. ChatGPT can be made silent and just listen, then respond when called by its name as a summon or activation word.

However, unlike Siri and Alexa, ChatGPT will continue to listen until the user manually turns it off.

OpenAI said users are automatically excluded from AI training when using voice mode. Audio clips are stored for 30 days so that the system has context from previous conversations, and can be deleted.

However, the increasingly human-like voice of AI also carries risks. CNET notes that there have been a number of cases and lawsuits related to the negative impact of humanizing AI, especially for people who are facing mental health problems.

OpenAI said the new model has expanded protections. The company also said the model performed better than previous models in key security areas, including self-harm, psychosis, violence, and sexual content.