A full-duplex voice model can answer in a fraction of a second. A web search or a database lookup takes several seconds. So the model hands the slow work to something else, keeps listening and talking while it runs, and speaks the result when it comes back. The designs differ in what does the slow work and who runs it.
Why the work is split
People leave about 200 milliseconds between turns. A search or a chain of tool calls takes seconds. A model that went quiet for that long would behave like a turn-based assistant again. Splitting the work keeps the voice responsive: one part holds the conversation, another does the slow work, and the result is put into the conversation when it is ready.
The model calls the tool and says nothing until the result is back.
The model says it is checking, answers a follow-up remark while the tool runs, then reports the result.
Three designs
In OpenAI's description of GPT-Live, the voice model handles the conversation and decides when to delegate. The backend holds the detailed instructions, business rules and tool workflows. The developer's application checks permissions, asks for confirmations and runs the functions that touch its own systems. Shizuku AI describes the same split for its Japanese model Shizuku-Duplex: an interaction model emits a DELEGATE token, keeps talking while a background model searches the web, and puts the result into its own words.
Read more: OpenAI, Google DeepMind, Shizuku AI
When the user interrupts
Interrupting the voice and cancelling the work are separate things, and the systems treat them differently. With GPT-Live, backend work can continue after the caller interrupts, and the application decides whether to finish or cancel it. In the Gemini Live API, an interruption discards the reply being generated and any pending function calls, and the server reports which calls were dropped.
For how a model decides that the user has really cut in, see Interruption and barge-in.
How it is tested
- Full-Duplex-Bench v3 plays 100 human-recorded requests, with hesitations and self-corrections, that need one to three chained tool calls. It scores tool-call accuracy, task success, latency and turn-taking. In the paper, the best system completed 60% of tasks at the first attempt, and every system failed on self-corrections and on multi-step reasoning in the hard scenarios. Its mock tools answer instantly, so it does not show what a model says during a slow tool call.
- τ-Voice runs agent tasks as live full-duplex calls under noisy, accented conditions. Agents that score about 85% in text drop to 26 to 38% once voice and real time are added.
- Shizuku AI ran its own Japanese benchmark, which includes looking things up on the web: Shizuku-Duplex answered 43% correctly, GPT-Live 1 50% and Gemini 3.8 Live 43%. The company that makes one of the models ran the test, so treat it as a starting point.
Guide: How full-duplex voice AI is evaluated
What to check
- Whether the voice keeps listening while a tool runs, or waits for the result.
- What the user hears while waiting: a spoken update, or silence.
- Who runs the backend, and whether you can choose the model or connect your own agent.
- What happens to work in progress when the user interrupts.
- How it is billed. OpenAI charges for the voice session and, separately, for the backend model and tools. Google bills one model by the token, with thinking included in the output price.
Models that call tools
The 14 current full-duplex models whose entry in the database says they call tools, grouped by how the tools run. Each model page gives the source.
| Model | How tools run | Access | Released |
|---|---|---|---|
Shizuku-DuplexShizuku AI | Hands the work to a backend model | — | 2026-11 |
Realtime-VenusAnt Group (inclusionAI) | Hands the work to a backend model | Open weights | 2026-09 |
DuplexOmniXi'an Jiaotong University / Peking University / Meituan | Hands the work to a backend model | Open weights | 2026-07 |
 GPT-Live 1OpenAI | Hands the work to a backend model | Hosted API | 2026-07 |
 Amazon Nova 2.5 SonicAmazon / AWS | Tool calls run in the background | Hosted API | 2026-10 |
 Gemini 3.8 LiveGoogle DeepMind | Tool calls run in the background | Hosted API | 2026-09 |
| Tool calls run in the background | Hosted API | 2026-07 | |
| Amazon Nova 2 SonicAmazon / AWS | Tool calls run in the background | Hosted API | 2025-12 |
| Calls tools | Hosted API | 2026-10 | |
Phonic-1Phonic | Calls tools | Hosted API | 2026-10 |
 StepAudio 3 RealtimeStepFun | Calls tools | Hosted API | 2026-09 |
| Calls tools | Open weights | 2026-08 | |
 Qwen-Audio Realtime (3.0 / 3.1)Alibaba Cloud | Calls tools | Hosted API | 2026-08 |
| Step-Audio 2 miniStepFun | Calls tools | Open weights | 2025-08 |
FAQ
- How does a full-duplex voice AI use web search or tools?
- It hands the slow work to something else and keeps the conversation going. The voice model decides that a search or a tool is needed, passes the request to a backend model or runs the tool call in the background, keeps listening and talking, and speaks the result when it comes back.
- Does the conversation stop while it searches?
- In a background design, no. The model says it is checking, keeps listening, and reports when the work is done. In an example Shizuku AI published, one web search took 7.1 seconds while the model kept talking. A tool call that blocks makes the voice wait for the result.
- Can I put my own agent or model behind the voice?
- Often, yes. GPT-Live 1 has a mode in which the developer connects an existing agent as the backend. In the Gemini Live API, a non-blocking function can forward the request to any agent or text model. DuplexOmni takes a pluggable thinking layer.
- Does a full-duplex model handle anything besides voice, such as a camera?
- Some do. Gemini 3.8 Live takes text, images and video frames along with audio. GPT-Live 1 takes no image or video input, so anything visual has to go through the backend. 15 models in the database are tagged video-in for camera or video input.
Updated