Open AI introduces GPT-Live

Open AI introduces a new generation of voice models capable of making conversations with artificial intelligence even more natural and human. We are talking about GPT-Live.

“A new generation of voice models for natural human–AI interaction, now powering ChatGPT Voice.”

The launch of GPT-Live

Open AI is launching a new generation of voice models, GPT-Live, its most intelligent voice model yet. Its strength, however, is that conversations now resemble real conversation more than ever.

This is possible thanks to a full-duplex architecture, which can listen and speak simultaneously. This differs greatly from previous approaches, making artificial intelligence much more “human”. As we speak, the voice model shows attention to what we are saying through typically human expressions such as “mm-hmm” or “yes”.

The resulting experience is completely different. Conversation becomes much smoother, and talking to models much easier.

At launch, GPT-Live will use GTP-5.5 in the background. At its global launch a couple of weeks ago, Open AI distributed two versions to users: GPT-Live 1 and GPT-Live 1 Mini.

Previous approaches

As we said, the main innovation is the conversational approach, which becomes as natural and human as possible. Collaborating with AI will feel almost like collaborating with another person.

To understand the significance of this change, we first need to examine earlier approaches.

Many language models use cascaded voice systems, based on models that act one after another, processing each individual turn. Cascaded systems produced slow, unnatural responses separated by long pauses.

Next came language models using turn-based voice systems. These reduced latency by generating audio within a single model. Although responses are faster and smoother, conversation still feels rigid and unnatural.

GPT-Live’s new approach

To create a completely different approach, Open AI examined and tried to solve the problems of earlier approaches.

To create continuous interaction and eliminate pauses caused by turn-taking, GPT-Live uses full-duplex architecture. This means the language model can process information, and even speak, while still listening. It continuously processes input even while generating output.

Conversation thus maintains a much more “human” sense of timing, producing more natural exchanges.

Another revolution also makes this possible: GPT-Live delegates more in-depth work. While GPT-Live manages uninterrupted interaction with the user, complex questions requiring research, reasoning and greater capabilities are delegated to another model, in this case ChatGPT-5.5. More complex tasks are therefore handled in the background while GPT-Live keeps the conversation going.

ChatGPT Voice

Finally, let’s discuss ChatGPT Voice. Around 150 million people interact with ChatGPT each week using features such as Voice and Dictation.

This lets users receive help without even using their hands to type. The language model is also used to practise languages, listen to stories and have simple conversations when we are bored.

Since GPT-Live’s introduction, conversations through the Voice button will therefore be more natural, with significantly better listening.