ChatGPT Voice Lets Users Interrupt, Revise Requests, and Resume Conversations

OpenAI describes ChatGPT Voice as “bidirectional” and “full duplex,” then translates that jargon into a simple interaction: “You can listen and talk at the same time.” In the demonstration, the assistant says a user can interrupt when it is “waffling on,” ask what it is talking about, and continue the exchange.

A word puzzle gives that claim a small, concrete test. Asked for an eight-letter word for something someone might dig up, the assistant suggests “artifacts,” “skeleton,” and then “potatoes.” When the user asks to pick the puzzle up later, it agrees: “Here when you need me.”

The exchange shows a pause and a resumption, not a test of how well the assistant remembers the puzzle after a long gap. OpenAI’s broader claim is that the new voice can keep up better than before; the scene’s narrower demonstration is that a user can stop the conversation without needing to conclude it formally.

A family trip to Casablanca supplies the more involved example. The user says the trip is for a 70th birthday and asks whether the food will suit a vegan. Before the assistant answers, the user asks for a restaurant list by email, then asks about Morocco’s weather in September.

The assistant responds to the requests together. It says the weather is lovely in September, names vegetable tagines and bean stews as vegan options, and says it has emailed a list of places. OpenAI says the voice can hold “multiple bits of information at once” and keep the conversation moving as plans change. In this exchange, the visible evidence is that the answer brings the food, restaurant, and weather requests into one response rather than handling them as isolated turns.

The demonstration does not show the email itself or establish how the assistant selected the restaurants. Its focus is the conversational flow: the user adds requests in succession, and the assistant replies after hearing the set.

The Spanish example moves from a language check into a message for a friend. The user asks whether “Buenos días” was right; the assistant corrects it to “buenas.” The user then asks it to say they will sell a guitar for no less than 150 pounds. The assistant translates the sentence as “Le venderé mi guitarra por no menos de 150 libras.”

The user adds, “It’s not old, it’s vintage,” and the assistant translates that too: “No es vieja, es vintage.” OpenAI calls this real-time translation. The exchange shows the task continuing through a correction and an added qualification, rather than ending with the first translated sentence.

OpenAI presents the voice as its most powerful model yet and says it listens while it speaks, keeps up better than before, and is “smarter than ever.” The scenes support a narrower description of what is being demonstrated: a user interrupts or pauses, adds requests before an answer, and revises a message during translation.

The video does not provide a broader evaluation of accuracy or reliability. In the travel scene, for example, the assistant gives a weather answer and says it emailed restaurants, but the demonstration does not show the email or the underlying information. The examples are product demonstrations of conversational features, not evidence about performance across other situations.

After the scripted exchanges, someone asks whether ChatGPT can pick up the grandchildren. The reply—“When it can make me a cup of tea, it will”—ends on the distinction the demonstrations have been drawing: the assistant can respond within a conversation, but the joke marks the limits of what it can do beyond one.