OpenAI said on Thursday it has improved ChatGPT Voice by applying a new speech model, GPT-Live, that understands conversational context in real time. The update reduces cut-offs and response delays and strengthens real-time interpretation.

Existing voice AI converted a user’s speech to text, then a large language model (LLM) generated an answer and played it back as speech. The process required a separate step to judge when the user had finished speaking, which sometimes caused conversations to break off or responses to be delayed.

OpenAI explained that GPT-Live continuously processes audio and identifies conversational context in real time. It can respond while a user is still speaking and handle a question that is paused and then resumed. It can start preparing an answer before the user stops speaking, speeding up responses.

OpenAI separated voice conversation functions from reasoning functions when designing ChatGPT Voice. When a user asks a question that needs additional information or requests a task that takes time, the AI keeps the conversation going while reflecting the answer as soon as it is ready.

ChatGPT Voice also improved its real-time interpretation. Previously it translated sequentially after an utterance ended, but the update continuously processes audio and interprets immediately in line with the flow of conversation. OpenAI stressed it can be used for meetings between users who speak different languages, travel, education and customer service.

OpenAI plans to expand the GPT-Live architecture across a range of backend models, input and output methods, and application programming interface (API) use cases. It will develop the technology to implement in voice conversations tasks that can be performed with text or AI agents.