The Transcript Had Second Thoughts

You say, “No—Tuesday, not Thursday,” just as the voice assistant begins to answer. It gives a quick mm-hmm, stops, and listens. On the screen, the transcript briefly assigns your correction to the assistant. Then the words move. A sentence breaks in two. Tuesday returns to the right speaker.

The conversation felt continuous. The record had to be assembled afterward.

That small gap is one of the most interesting details in OpenAI’s new engineering account of GPT-Live. The system is full-duplex: it can listen and speak at the same time instead of waiting for a clean silence between turns. Deeper reasoning and tool use happen on a separate path, so the voice can keep the conversation moving while other work runs in the background.

This makes the exchange feel less like walkie-talkies and more like speech. People interrupt, correct themselves, offer tiny acknowledgements, pause without surrendering the floor, and occasionally begin the same sentence together. The model can respond to that rhythm without first chopping every contribution into a finished audio parcel.

The rest of the software still wants parcels.

Chat interfaces, analytics, safety systems, and tool calls are generally organized around discrete user and assistant messages. OpenAI’s application server therefore has to infer those messages from the continuous stream. The newest message remains provisional. Its text, timing, and speaker assignment can change as more audio arrives. Only after one speaker has held the floor long enough does the system finalize the turn.

Internally, the product keeps two views: a speculative conversation state that can update on screen and an authoritative transcript for the systems that need a stable record.

This is good engineering. It is also a useful confession about transcripts.

A transcript looks like evidence because it has timestamps, names, and sentences wearing sensible punctuation. In live conversation, those details are interpretations. Someone has to decide whether “right” was an acknowledgement or the beginning of a reply, whether a correction replaces the earlier phrase, and whether an interruption belongs inside one turn or creates another. Full-duplex speech does not cause this ambiguity. It merely makes the software admit that the ambiguity was always there.

Current systems are not finished with the problem. Full-Duplex-Bench-v2, a recent multi-turn evaluation, found that full-duplex agents still become confused when speakers overlap, handle corrections poorly, and lose track of who or what is being discussed. Conversational smoothness can improve while record quality remains unsettled.

That distinction matters once voice becomes more than a way to ask for weather. GPT-Live can decide whether to interrupt, continue listening, or invoke a tool many times per second. It can delegate search and deeper reasoning while the person keeps talking. The transcript is no longer only a souvenir of the conversation. It may become the boundary between what someone seemed to say and what the system understood as an instruction.

Imagine a support call where “don’t cancel it yet” is briefly split into a command and a correction. Imagine a worker thinking aloud about two customer records while an assistant prepares an update. The important interface question is not whether speech recognition eventually produces a plausible paragraph. It is which interpretation was allowed to act, when it became stable, and whether the person could still correct it.

We have argued that AI meetings need a guest list: people should know who is listening, what will be kept, and when a room becomes a record. Full-duplex voice adds a more mechanical question. At what moment does the record stop being a draft?

The answer cannot be “whenever the backend needs one.”

Live text is already an accessibility surface. W3C guidance on captions and transcripts notes that live captions make spoken material available as it happens, while a recording posted afterward will probably need editing for accuracy. A provisional transcript cannot be treated as disposable interface glitter when someone may depend on it to participate in the conversation. The live view has to be readable and timely. The durable view has to be correctable and honest about what changed.

That suggests a modest set of transcript manners. Mark the current turn as live without making it flicker like a stock ticker. Keep corrections visible long enough for the person to notice them. Distinguish a listening acknowledgement from a substantive answer. When speech triggers an action, preserve the exact interpretation that crossed the boundary and give the speaker a direct way to contest it. If the final record differs materially from what appeared live, let that revision leave a small trace.

Nobody needs version control for every mm-hmm. A natural conversation would become unbearable if each syllable arrived with a compliance badge. The design work is to identify where fluidity becomes consequence.

This extends the case that work no longer fits inside a conversation. Agent runs need draft marks around changes in evidence, scope, and state. Full-duplex voice shows that the conversation itself may need one too. The words can look finished before the system has finished deciding who said them.

There is something humane in allowing the transcript to have second thoughts. People have them constantly. We revise mid-sentence, speak over one another, and use tone to rescue words that look harsher on the page. A system that quietly reconstructs all of that into an immaculate history risks making conversation seem more settled than it was.

The better interface lets the transcript remain malleable for a moment. Then, before it becomes memory, evidence, or instruction, it lets the people who spoke see the form that is about to set.