OpenAI has unveiled a new family of voice models called GPT-Live, designed to simultaneously process user speech and generate spoken responses. According to the company, this makes interactions with ChatGPT feel more natural, as the system can now react to pauses, changes in speech tempo, and interruptions during a conversation.
OpenAI noted that the first version of ChatGPT’s voice mode, launched in 2023, relied on a pipeline of separate models: speech was first converted to text, then a response was generated, and finally it was spoken aloud. This approach introduced delays and made the dialogue feel less fluid.
Later, the Advanced Voice mode was introduced, where speech processing and response generation were handled by a single multimodal model. However, conversations still followed a turn-based structure: the system would only begin replying after the user had finished their statement.
The new GPT-Live model employs a full-duplex architecture, allowing it to analyze incoming audio streams and produce its own speech at the same time. The system continuously determines when to speak, when to keep listening, when to ignore brief pauses, or when to respond to interruptions. During a conversation, the model can use short affirmations, creating the impression of a real-time dialogue.
OpenAI also reported that GPT-Live supports real-time synchronous translation, better isolates the user’s voice from background noise, and can display informational cards with weather, stock market data, or sports scores. For complex queries, the voice model hands off the task to more powerful language models within the company, while maintaining the conversation until a response is ready.
According to OpenAI, the full-featured GPT-Live-1 is available to subscribers of ChatGPT’s paid plans, while the lighter GPT-Live-1 mini is accessible to users of free accounts, as reported by N+1.
About the Author
