Is a cascade with barge-in full-duplex?

It behaves like one when you interrupt it, and many products describe it that way. The database labels it Cascade, because speech recognition, the language model and speech synthesis still run one after another.

Which models are full-duplex?

37 of the 70 models in the database. Filter Models by Full-duplex to see them.

Where can I find full-duplex training data?

31 of the 107 datasets are recorded with one channel per speaker. Filter Datasets by Full-duplex, then check the licence and source of each.

Why do response times differ between pages?

Vendor figures and benchmark runs measure different things under different conditions. The database names the source next to every number.

Can I use these datasets commercially?

It depends on each licence. Use the Commercial OK filter on Datasets, and read the original terms before you ship anything.

What is full-duplex voice AI?

Full-duplex voice AI is a speech system that listens and speaks at the same time, like two people on a phone call. It keeps listening while it talks, so you can interrupt it, it can say "mm-hm" while you speak, and it recovers when you both start at once. Half-duplex systems take turns, like a walkie-talkie.

Guides

Turn detectionHow a voice agent decides that the user has finished, and how end-of-turn models compare on TurnBench and eot-bench.How full-duplex voice AI is evaluatedWhat the benchmarks measure, how Full-Duplex-Bench v1 to v3 differ, and why scores change between runs.

What full-duplex means

The terms come from telecommunications. A half-duplex channel carries signal one way at a time, like a walkie-talkie: push to talk, release to listen. A full-duplex channel carries signal both ways at once, like a telephone call. In networking the word describes links such as Ethernet. In voice AI it describes a model that keeps listening while it talks.

Why it matters

In human conversation, the most common gap between one person finishing and the next person starting is about 200 milliseconds. Stivers and colleagues found that figure across the ten languages they measured in 2009. Planning a single spoken word takes at least 600 milliseconds, so people start drafting a reply while the other person is still talking. A listener who waited for silence before planning would answer 800 to 2,000 milliseconds later. In spontaneous speech corpora, overlaps make up roughly 40 percent of changes between speakers.

Four behaviours

Barge-inYou cut in while the system is talking, and it stops and listens.

BackchannelThe system says short acknowledgements such as "mm-hm" while you talk, without taking the turn.

Overlap and recoveryYou both start at once, one of you yields, and the conversation recovers within a second.

Co-completionThe listener supplies the word the speaker is looking for. No public benchmark measures it directly yet.

Guide: Turn detection, how a voice agent decides that the user has finished.

How systems are built

The database labels every model by how it handles the conversation loop. The timelines show the same exchange with each kind of system.

Half-duplex

Waits for you to stop, then answers. What you say while it talks is not heard.

Cascade

Speech recognition, a language model and speech synthesis run in turn, which adds a gap before it answers. Many stop talking when you cut in.

Full-duplex

One model listens while it speaks. It can say “mm-hm” while you talk, answer after a short gap, and give way when you cut in.

Your speechAI speechProcessingNot heard or cut offTime →

Full-duplexOne model listens and speaks at the same time. Examples: Moshi, gpt-realtime, Gemini 3.8 Live.37 models

CascadeSpeech recognition, a language model and speech synthesis run in sequence. Many cascades stop speaking when the user cuts in, so they handle barge-in, but each stage waits for the one before it. Examples: LiveKit Agents, ElevenLabs Agents, Vapi.20 models

Half-duplexThe system takes turns. It listens until you stop, then speaks.6 models

Full-duplex describes how the conversation flows, so a well-built cascade can come close to it. On this site the Full-duplex label is kept for single models that listen and speak at once. Cascades are labelled Cascade even when they handle interruptions.

How it is measured

Full-Duplex-Bench (Lin et al., 2025) scores pause handling, backchanneling, turn-taking and interruption. Much of the scoring rests on one rule: the model has taken the floor if it says more than three words or speaks for more than a second. When the user has finished, that rule should fire. When the user has only paused, it should not. Later versions add real disfluencies and tool use.

The database reports response time in two ways: the vendor's own first-response figure, and Full-Duplex-Bench runs where they exist. They measure different things, so compare like with like.

Guide: How full-duplex voice AI is evaluated. Read more: Mapping the benchmark landscape. In the database: 27 full-duplex benchmarks, and Analytics for hosted APIs.

What data it takes

A full-duplex model needs conversations recorded with each speaker on their own channel. Separate tracks keep overlaps, interruptions and backchannels intact. On a single mixed track they blur together, and diarization has the most trouble exactly where speakers overlap. The database marks these corpora Full-duplex. Their licences and consent terms vary, so check the Licence and Source columns before training.

Read more: The data ceiling. In the database: 31 full-duplex datasets. The whole catalogue is also a Hugging Face dataset (CC BY 4.0).

Glossary

Backchannel: A short listener acknowledgement, such as "uh-huh" or "right", produced while the other person is still talking.
Barge-in: Interrupting a system while it is speaking. A full-duplex system stops and listens.
Cascade: Speech-to-text, a language model and text-to-speech wired in series.
Channel-separated: A recording with each speaker on a separate track.
Full-duplex: Telephone-style conversation, where both sides can listen and speak at once.
Full-Duplex-Bench (FDB): A family of benchmarks, v1 to v3, that measure full-duplex conversational behaviour.
Half-duplex: Walkie-talkie-style conversation, one direction at a time.
Neural audio codec: A model that turns a waveform into discrete tokens a language model can read and write, and turns them back into sound.
Speech-to-speech (STS) model: A model that takes audio in and produces audio out without a text-only stage in between.
Speech language model: A model that predicts speech as a sequence of tokens, the way a text model predicts words.
Time to first audio (TTFA): How long a system takes to start producing sound after it receives input.
Turn-taking: How speakers settle who talks next, and when.

FAQ

Is a cascade with barge-in full-duplex?: It behaves like one when you interrupt it, and many products describe it that way. The database labels it Cascade, because speech recognition, the language model and speech synthesis still run one after another.
Which models are full-duplex?: 37 of the 70 models in the database. Filter Models by Full-duplex to see them.
Where can I find full-duplex training data?: 31 of the 107 datasets are recorded with one channel per speaker. Filter Datasets by Full-duplex, then check the licence and source of each.
Why do response times differ between pages?: Vendor figures and benchmark runs measure different things under different conditions. The database names the source next to every number.
Can I use these datasets commercially?: It depends on each licence. Use the Commercial OK filter on Datasets, and read the original terms before you ship anything.

Cite this page

@misc{fullduplex_knowledge,
  title        = {What is full-duplex voice AI?},
  author       = {{Fullduplex}},
  year         = {2026},
  howpublished = {\url{https://www.fullduplex.ai/knowledge}}
}