A half-duplex voice system takes turns: it listens until you stop, then speaks. A cascade runs speech recognition, a language model and speech synthesis one after another, and many stop when you cut in. A full-duplex model is one speech-to-speech model that listens while it speaks, so it can answer quickly, give way and say "mm-hm".
The same exchange, three ways
The timelines play one exchange with each kind of system. Press Interrupt while the AI is answering to see what each one does when you cut in.
Keeps talking. Your interruption is not heard.
Stops after a processing delay, then answers after another pause.
Stops almost at once and answers after a short gap.
Half-duplex: Keeps talking. Your interruption is not heard. Cascade: Stops after a processing delay, then answers after another pause. Full-duplex: Stops almost at once and answers after a short gap.
The terms come from telecommunications. A walkie-talkie carries one direction at a time, and a telephone carries both at once. On this site the Full-duplex label is kept for single models that listen and speak at once. A cascade that stops when you cut in is still labelled Cascade, because its parts still run one after another.
Measured results: Interruption and barge-in, Backchannels
Side by side
| Half-duplex | Cascade | Full-duplex | |
|---|---|---|---|
| Response time | Waits for a stretch of silence, then starts its reply. | Each stage waits for the one before it. A well-tuned cascade lands near one second end to end, and fast parts can bring it under a second. | Around 200 ms in the best reported measurements. The Moshi paper gives 160 ms in theory and 200 ms measured on an NVIDIA L4. |
| When you cut in | Keeps talking to the end of its reply. What you say meanwhile is not heard. | Many stop once a voice activity or interruption detector hears you. | Hears you while it talks, so it can stop mid-word and respond. |
| Backchannels | None. It makes no sound while it listens. | None by default, because the reply starts after your turn ends. A separate backchannel model can add them. | Can say "mm-hm" while you talk. In a 2026 benchmark, most models almost never did so when one was due. |
| What the model hears | Varies. Some half-duplex models take audio in directly. | A transcript. Pitch, emotion and pace are dropped when speech becomes text. | The audio itself, including how something was said. |
| Building one | The simplest loop: one rule decides that you have finished, often a silence timeout. | Assembled from speech recognition, language model and speech synthesis services. Each stage's output can be inspected, which makes it easier to debug. | Use a hosted API or open weights. Training one takes conversations recorded with each speaker on their own channel. |
| Swapping parts | Varies by system. | Any stage can be replaced with a different recognizer, language model or voice, so a team can keep parts it already runs. | One model listens and speaks, so you choose the whole model rather than its parts. |
Latency figures from different kinds of system measure different things, so compare like with like. The backchannel result is from DuplexAct, covered in Signals 2026-W43. Read more: From pipeline to integrated, Speech-to-speech AI, a primer
Cascaded vs end-to-end
A cascade turns speech into text, works on the text and turns the answer back into speech. An end-to-end, or speech-to-speech, model takes audio in and produces audio out without a text-only stage in between. The transcript keeps the words but drops how they were said. A recognition error early in a cascade cannot be fixed later, because no later stage hears the original audio.
Cascades are easier to debug, and in narrow domains where accuracy matters most, a cascade with tuned speech recognition often still does better on the task.
Not every speech-to-speech model is full-duplex. Some take audio in and give audio out but still take turns, and the database labels them Half-duplex. Other designs sit in between, such as a cascade with a small model on top that decides at short intervals when to speak, listen or stop. From pipeline to integrated sorts these designs into four families.
Releases by quarter
The 63 conversation models in the database by quarter of release, stacked by how each one handles a conversation. Select a quarter to list its models. The 7 models labelled Other, such as text-to-speech, are left out.
Full-duplex: Amazon Nova 2.5 Sonic, Griffin-Lite, Grok Voice 3.0 (xAI Speech to Speech API), Phonic-1
Cascade: Cartesia Managed Agents, Eleven Realtime (v1 Mini / v1 Max), Mistral Voice Agents (Voxtral 3)
Half-duplex: Apriel Voice 27B Preview, LFM3-Audio-2B / LFM3-Omni-2B, Rime Prism
27 of the full-duplex models in the database were released in 2026, against 8 in 2025. The chart counts what the database lists, so older releases may be missing. See every model on Models.
Which to use
OpenAI's developer documentation describes two valid tracks for voice: chained pipelines, which stay reliable and are easier to debug, and speech-to-speech models, which aim for lower latency and more natural conversation.
A cascade suits narrow tasks where accuracy matters most, teams that already run their own recognition, language and voice parts, and products that need to inspect each step. A team with strong parts can also put a turn-taking model on top of its cascade and get some full-duplex behaviour without retraining the whole stack. A full-duplex model suits open conversation, where interruptions, backchannels and fast replies matter.
Whichever you pick, test interruptions and backchannels on your own calls. A high turn-taking score says little about either.
Guides: Turn detection, Interruption and barge-in, Backchannels, Evaluation
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.
- Is speech-to-speech the same as full-duplex?
- No. Speech-to-speech means audio in and audio out with no text-only stage in between. Full-duplex means listening while speaking. Moshi is both, but some speech-to-speech models still take turns, and the database labels those Half-duplex.
- Which is faster, a cascade or a full-duplex model?
- In the best reported measurements, full-duplex models answer in around 200 ms, against around one second for a well-tuned cascade. Published figures measure different things, so check how each one was measured before comparing.
- How many full-duplex models are there?
- 37 of the 70 models in the database are full-duplex, 20 are cascades and 6 are half-duplex. The rest are other kinds of speech model, such as text-to-speech.
Updated