Explainer
What Does Full-Duplex Mean? Dessa AI's Key Idea in Plain English
"Full-duplex" is the phrase Block uses most often for Dessa — and the phrase most readers nod at without quite knowing what it changes. This is the plain-English version: no engineering degree required, just the difference between a walkie-talkie and a phone call, and why that difference decides whether talking to an AI feels like issuing commands or having a conversation.
The one-sentence definition
Full-duplex means both sides can send and receive at the same time. In voice AI, it means the system can listen while it is speaking, and speak while it is still listening to you. That is it. Everything else — the laughter, the interruptions, the natural timing in Block's Dessa demos — flows from that one ability.
Walkie-talkie versus phone call
A walkie-talkie is half-duplex. Only one person can transmit at a time. You press, talk, say "over", release, and wait. If both people press at once, neither is heard properly. Much of today's voice AI works like that, even when it does not feel quite so rigid: you speak, there is silence while the system transcribes, thinks and generates, then it speaks, and while it speaks it is mostly deaf to you until it finishes or you barge in with a wake action.
A normal phone call is full-duplex. You can say "mm-hmm" while the other person is mid-story. They can hear you laugh at the funny line without stopping. If you both start talking at once, you both notice instantly and one of you yields. Nobody says "over". Human conversation has always been full-duplex — which is why half-duplex AI feels subtly exhausting. You are doing phone-call behaviour into a walkie-talkie system.
The small sounds that carry a conversation
Once a system can listen and talk simultaneously, a whole layer of conversation becomes possible. Linguists call the listener's short responses backchannels: "yeah", "got it", "ooh", "right". They are not interruptions; they are proof of attention, and they tell the speaker to keep going, slow down or clarify. Block says Dessa backchannels in exactly this way, laughs the moment something is funny rather than after a polite pause, sighs, hears sarcasm and emotion, and knows when to cut in and when to let you finish. We explain the underlying speech-only design in how Dessa AI works.
Interruption is the clearest test. In a half-duplex system, interrupting is an error state: stop the audio, discard the plan, start again. In a full-duplex system, interruption is information. A person who cuts in with "wait — Tuesday, not Thursday" is correcting the conversation in real time, and a system that can hear that while speaking can adjust instead of finishing the wrong sentence beautifully.
Why text in the middle breaks duplex
Traditional voice assistants are a relay: speech to text, text model, text to speech. Each handoff needs a reasonably complete chunk before the next stage can work, which pushes the system toward turns. Worse, the transcript strips out exactly the signals duplex conversation runs on: timing, tone, hesitation, a laugh inside a sentence, a voice cracking. Block's blunt summary is that interacting with current voice AI "can just plain suck" because the most powerful AIs cannot hear how you feel.
Dessa removes the relay. Trained from scratch on human conversation audio only — no text, no pretrained models — it takes audio in and produces audio out directly. Block calls it the world's first realistic live-conversation model trained only from human speech, and describes inference faster than real time running locally on a MacBook Pro with an M4 Max chip or better. Whether that makes it better overall than text-backed rivals is a separate question we tackled in Dessa vs ChatGPT Voice; on timing and conversational feel, duplex is the structural advantage.
Full-duplex does not mean perfect
Three cautions keep this honest. First, full-duplex is a capability, not a quality score: a system can listen while speaking and still answer badly. Block has published no benchmarks, parameter count or independent evaluation for Dessa, so natural demos are evidence of conversational timing, not of factual reliability. Second, Block's demos are its own live, unsteered recordings — impressive, but selected by Block, and reviewers noted occasional awkwardness, including Dessa calling a speaker "Ronaldo". Third, Dessa is a research preview: no public release date, pricing or API has been announced, and early access goes first to a small group of AI safety researchers.
How to spot full-duplex in the wild
When you test any voice AI, try four things. Talk over it mid-sentence and see whether it notices gracefully or ploughs on. Laugh or sigh and see whether it reacts in the moment. Say "yeah" quietly while it talks and see whether it treats that as encouragement rather than a new command. Correct a detail mid-answer — "no, the other Lisbon" — and see whether the reply changes course. A walkie-talkie system fails these politely. A full-duplex system handles them the way a person does: imperfectly, instantly, and without breaking the flow.
That is why Block chose the term. "Full-duplex" sounds like a telecom spec, but for Dessa it is the whole product idea in one word: conversation is not turns, it is overlap — and a model that cannot overlap cannot quite talk with you, only at you, in turns.
Sources: Block's official Dessa page (block.xyz/dessa) for full-duplex behaviour, backchannelling and local inference claims; Kyutai's Moshi paper (arXiv:2410.00037) for the full-duplex comparison; The Verge and TechCrunch, October 2026.
Frequently asked questions
What does full-duplex mean in simple terms?
It means both sides can listen and speak at the same time, like a phone call instead of a walkie-talkie. A full-duplex voice AI can hear you while it is talking, so it can react, backchannel and handle interruptions in real time.
What is the difference between half-duplex and full-duplex voice AI?
Half-duplex systems take strict turns: you speak, then the AI speaks, and it largely cannot hear you while it talks. Full-duplex systems process incoming and outgoing speech simultaneously, making timing, laughter and corrections feel natural.
Is Dessa AI full-duplex?
Yes, according to Block. Dessa is described as a full-duplex, speech-to-speech model trained only on human conversation audio, able to backchannel, laugh mid-sentence and handle interruptions. It is a research preview with no public release date announced.
🚀 Get Dessa launch alerts
Dessa is still a research preview. Join the free DessaAI Daily list and we will email you the day access opens — plus the week's biggest Dessa stories, once a week. No spam, no hype.
Form not loading? Subscribe on the signup page →