GPT-Live Combines Full-Duplex Voice With Backend Delegation
OpenAI developer Aarti Bagul describes GPT-Live as combining full-duplex voice interaction with the ability to delegate complex tasks to a backend model or system. The first change lets the model listen while speaking, rather than waiting for a user’s turn to end; the second separates the voice exchange from the work behind it. Bagul says developers can use prompts and delegation controls to manage interruptions, collect required details before handing off a task, and keep critical application data out of conversational context alone.
Full duplex changes turn-taking; delegation separates conversation from complex work
? aarti-bagul describes GPT-Live through two distinct changes: it can listen and speak at the same time, and it can hand complex work to a backend model or system. The first changes how a voice interaction handles turn-taking; the second changes where work gets done.
Earlier voice systems chained together speech-to-text, a language model to reason and respond, and text-to-speech. Each handoff added latency, making responses feel slow and stilted. GPT Realtime brought audio processing, response generation, reasoning, and tool decisions into one model. That made speech-to-speech interaction faster and more natural, but it remained primarily turn-based: the model had to decide when a user had finished speaking before responding. That can make it difficult to distinguish a brief acknowledgment from an interruption or correction.
GPT-Live’s full-duplex design lets it listen while speaking. Bagul says this helps it distinguish whether someone is following along or trying to interrupt, making the exchange feel more natural.
Delegation is a separate change. GPT-Live manages the voice interaction and can pass harder tasks to a backend model or system. Developers can update that backend as newer models become available, or connect GPT-Live to an existing text-based application to add a voice interface.
Background noise is a question of conversational intent
In a busy environment, the challenge is not just detecting speech. Turn-based systems can be confused by background noise because they are trying to determine whether someone is speaking at all. GPT-Live can also use the conversation’s context to decide whether to keep speaking, stop, or interject.
Developers can guide that behavior in the prompt. For example, they can tell the agent to ignore side conversations and respond only when addressed. In that setup, GPT-Live can listen to a conversation without chiming in unless it is spoken to.
Examples span transactions, support, lessons, and software
? aarti-bagul points to four applications: Yelp uses GPT-Live for restaurant reservations and food orders; Fin uses it for customer support; Speak uses it for live language lessons; and Cognition has added voice to its AI software engineer, Devin.
For a French voice agent, Bagul recommends first trying the newer language and regional voice options. She plays two French voices, Brise and Sillage, and notes that several new voices, including those two French options, had been released since the developer’s question was posted.
Developers can also make a custom voice in two ways: describe its accent, tone, and pacing in natural language, or generate it from a recording—with the speaker’s consent. Prompting can further shape the voice’s pacing and personality, its level of formality, and the pronunciation of particular words.
Reliable delegation depends on separating the two prompts’ jobs
A developer reports that GPT-Live sometimes says it will pass a request along but no backend request arrives. ? aarti-bagul advises keeping GPT-Live’s prompt short and focused on when to delegate and when not to. The backend prompt should hold the details of which tools to use and the operating procedure. As Bagul puts it, “The model only needs to know when to delegate and when not to delegate.”
Bagul recommends being explicit about the delegation policy. She says OpenAI had updated its GPT-Live 1 prompting guidance and that GPT-Live 1 tends to perform better with the special <|SpawnThinking|> token. This advice applies specifically to GPT-Live 1; developers should check the current guidance for the model they are using in the GPT-Live prompting guide.
For an additional safeguard, Bagul describes session.triggers, a native integration with the Decisions API. A trigger can evaluate a predefined question, such as whether to delegate. If the API indicates that delegation is needed and GPT-Live has not initiated it, the application can force the delegation.
Collect required details before sending a request to the backend
The opposite problem is over-delegation: a voice agent passes work to the backend before it has collected the information needed to complete it. For a booking flow requiring a name, date, time, and address, ? aarti-bagul recommends client delegation to control the back-and-forth between GPT-Live and the backend.
The backend can send an update that gives GPT-Live its immediate next question—for example, “ask the user for their name”—through session.commentary.append. A separate thinking update, sent through session.thinking.append, can tell the model what else it must collect: after the name, ask for the date, time, and address, and do not delegate again until those details are gathered. The instruction can also require GPT-Live to confirm the fields with the user before delegating.
Persist critical details in the application, not only in conversation context
? aarti-bagul says compaction of the voice model’s context is handled automatically. But if information is critical to the application, she recommends saving it there: compaction might lose key details.
She distinguishes two delegation paths. Built-in delegation through Responses or Decisions automatically gathers conversation context and passes it to the backend. Client delegation gives the developer more control over which context goes to the backend and which updates return to GPT-Live.