Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
GPT-Live is a full-duplex voice API for building agents that listen and speak at the same time, rather than waiting for one party to finish before the other responds.
Full-duplex means the audio channel carries data in both directions at once, similar to a telephone call: the agent keeps listening while it's speaking, and the user can talk without waiting for a pause. This is in contrast to half-duplex systems, like a walkie-talkie or turn-based speech-to-speech models such as the GPT Realtime API, where only one party can transmit at a time and the other must wait to respond.
GPT-Live marks a shift away from turn-based voice interactions toward natural, bidirectional, phone-call-like conversations. It enables more natural, real-time communication that feels as natural as talking to another person.
Why use GPT-Live?
The turn-based pattern of earlier speech-to-speech models can feel unnatural, especially when a user wants to interject, acknowledge, or redirect the conversation mid-response. GPT-Live's goal is a more natural, bidirectional, phone-call-like conversation instead of a turn-based interaction. See Capabilities for details.
How GPT-Live works
GPT-Live splits a voice application into two parts:
- The live voice model handles the spoken conversation. It listens, speaks, decides when to respond, and decides when to hand off work. You steer this behavior with a short set of session instructions that cover conversational style and when to delegate.
- A backend handles delegated reasoning, lookups, tool selection, and longer-running tasks. Hosted tools can run through a configured Responses API backend, while your application executes private functions and remains responsible for permissions, user confirmations, business records, and durable task state.
Your application owns everything outside the spoken exchange: permissions, confirmations, tool execution, business records, and durable task state. Because the live model can keep talking while backend work runs, interrupting speech doesn't cancel that work. Your application decides whether to finish, change, or cancel it.
For guidance on splitting instructions between the voice model and the backend, see Delegate work in GPT-Live.
Capabilities
| Capability | Description |
|---|---|
| Full-duplex conversation | Listens and generates a response at the same time, without rigid turn-taking. |
| Natural interruptions | Adapts when a user interrupts, instead of finishing a queued response. |
| Delegation | Hands off requests that need search, tool use, or deeper reasoning to a configured backend model or to your own application, while the live session continues. See Delegate work in GPT-Live. |
| Speech translation | Translates speech across source and target languages and locales. Quality varies by language pair. |
| Output | Voice only. |
Supported models
| Model | Description |
|---|---|
gpt-live-1 |
The flagship full-duplex voice model. |
Language and translation support
gpt-live-1supports multilingual speech and can be used for speech-translation experiences. Quality varies by language and locale, with no guarantee of uniform quality or parity with gpt-realtime-*.
Quality varies by language pair, with no guarantee of parity with gpt-realtime-translation.
Connect to GPT-Live
GPT-Live supports three transports:
- WebSocket: A trusted backend or middle-tier service connects directly and streams audio as base64-encoded raw audio using 16- or 24-kHz PCM, or 8-kHz G.711 μ-law (PCMU) or A-law (PCMA). See Use GPT-Live for real-time voice.
- WebRTC: Browser or native clients connect with low-latency, negotiated media tracks. See Use GPT-Live via WebRTC.
Choose a voice architecture
GPT-Live is one of several ways to build a voice experience. Choose based on how speech connects to reasoning and tools.
| Architecture | Best for | Why choose it |
|---|---|---|
| GPT-Live | Full-duplex conversations with a separate backend | You choose the voice model and the backend that reasons and uses tools independently, and the conversation continues while backend work runs. |
| GPT Realtime API | Speech, reasoning, and tool use in one turn-based model | One model interprets audio, decides what to do, and responds in speech. |
| Chained pipeline | Full control over each speech and text stage | Speech-to-text, your own agent, and text-to-speech run as separate stages that you can inspect or replace. |
Context management in long conversations
GPT-Live manages conversation context automatically as a session grows; you don't set a parameter to enable it. The instructions you provide at session start are preserved for the life of the session.
As the conversation approaches the context window limit, GPT-Live summarizes older history in the background and continues within the same session. Older details can be summarized or dropped, so keep important facts, confirmed actions, and current task state in your application and supply relevant context when it's needed.
Speech and task work run independently
Because GPT-Live is full duplex, the spoken conversation and delegated backend work proceed on separate tracks:
- Interrupting the assistant's speech stops it from talking, but doesn't cancel work already running in the backend. "Stop talking" and "Cancel my order" are different intents.
- A completed backend response doesn't mean the user heard the answer. Verify spoken confirmations against your application's authoritative state.
- Transcript fragments reflect audio cadence, not complete turns. A fragment isn't a finished user turn, user and assistant text can overlap, and transcripts can contain mistakes.
Cost model
GPT-Live bills the voice conversation separately from the backend that reasons and uses tools:
- The voice session covers the spoken conversation, including silence and time when the backend is working.
- The backend model and tool usage is billed the same as an application without voice.
Session usage is reported as a cumulative running total, so read the latest usage value rather than adding snapshots together. For current rates, see the Azure OpenAI pricing page.
Limits
The GPT-Live API limits the number of concurrent sessions per subscription based on your tier. For the per-tier values, see GPT-Live concurrent session limits.