en.Wedoany.com Reported - US-based artificial intelligence company OpenAI has launched a new generation voice model, GPT-Live, designed to power more natural human-machine voice interactions. Based on a full-duplex architecture, the model can continuously listen and respond in real time, moving away from the traditional turn-based pattern of voice assistants where "the user finishes a sentence, then the system starts to reply."
GPT-Live will be available in two versions: GPT-Live-1 and GPT-Live-1 mini. GPT-Live-1 will become the default voice model for ChatGPT Voice for Go, Plus, and Pro users, while GPT-Live-1 mini will serve as the default for Free users. OpenAI stated that GPT-Live is rolling out globally to ChatGPT users, covering iOS, Android, and ChatGPT.com.
Full-duplex capability is the core of this model update. Traditional voice interactions typically require waiting for the user to pause or finish speaking before deciding whether to respond, often leading to issues like interruptions, missed listening, delays, or unnatural conversational flow. GPT-Live can continuously monitor voice input during a conversation while determining whether the user is pausing, adding information, correcting themselves, or expecting an immediate response. The model handles pauses, interruptions, changes in speaking speed, and shifts in conversational rhythm, making voice interactions closer to real-time human communication.
GPT-Live is designed for real-time voice agent scenarios. The model must simultaneously handle voice understanding, context tracking, response generation, and voice output control, while also deciding when to continue listening, when to interject, and when to wait for the user to elaborate. Compared to the simple process of converting speech to text and then generating a response, the full-duplex voice model integrates listening, thinking, and speaking into a single continuous interaction system, placing higher demands on latency control, voice activity detection, context retention, and turn-taking judgment.
OpenAI has also released the GPT-Live system card, explaining that GPT-Live-1 and GPT-Live-1 mini belong to a new generation of voice models aimed at making AI conversations more natural and intelligent. The system card notes that full-duplex capability allows the model to continuously listen and respond, rather than waiting for a clear turn to end before acting.
In terms of product form, GPT-Live first enters ChatGPT Voice. Users can engage in voice conversations with ChatGPT on mobile or web platforms, with the model responding in real time based on voice input content and conversational context. For applications such as voice assistants, customer service bots, educational tutors, meeting assistants, in-car interactions, and real-time companionship apps, key metrics for this type of model will focus on response latency, interruption handling, natural pause recognition, continuous multi-turn dialogue, and voice output stability.
On the API front, OpenAI has opened the registration portal for GPT-Live-1 API notifications, allowing developers to apply for subsequent availability updates. If GPT-Live later enters the API, businesses and developers will be able to integrate full-duplex voice capabilities into voice customer service, smart hardware, real-time translation, remote collaboration, and industry-specific voice agent products.
