Explainer

Dessa AI Latency and Interruption: How It Knows When to Speak

2026-11-13 · 7 min read · By DessaAI Daily

Ask what makes Dessa feel different in Block's recordings and the answer is rarely a fact it knows. It is timing. Dessa laughs mid-sentence, answers a cut-in without ploughing on, drops in a “got it” while the other person is still talking, and gets louder the moment it is asked to speak up. This article takes that timing apart honestly: what Block has claimed about latency and interruption, what it has pointedly not measured in public, and why a speech-only model can do timing tricks that a text pipeline structurally cannot. The underlying idea is set out in how Dessa AI works; this piece stays on the clock.

Latency: the claim, and the missing number

Start with the gap, because it is the part most coverage skips. Block has published no latency figure for Dessa. No milliseconds. No end-to-end delay, no time-to-first-audio, no measurement conditions, no benchmark table at all. What Block says is qualitative: inference runs faster than real time, and it runs locally on a MacBook Pro M4 Max or better, “so there's no awkward wait”. Faster than real time is a real claim — it means the model can generate speech at least as fast as speech is spoken, which is a precondition for live conversation — but it is not a latency measurement, and it should not be quoted as one. Our benchmarks explainer lists everything else in the same unpublished category: model size, training hours, and any independent evaluation. If a site gives you a precise Dessa latency number, it did not get it from Block.

Interruption: listening while speaking

The technical term is full-duplex, and our plain-English guide to full-duplex covers it in full. The short version: a full-duplex system keeps its ears open while its mouth is moving. That sounds trivial — people do it constantly — but most voice AI does not. The standard pipeline waits for silence, transcribes, thinks in text, then speaks, and while it speaks it is effectively deaf. Interrupt it and you either talk over a recording or trigger a clumsy stop. Block says Dessa “knows when to cut in” and, by the same design, when it is being cut in on: because listening never stops, a new voice, a laugh or a frustrated sigh arrives as information the model can react to mid-sentence, not as an error after the fact.

Backchannels: the smallest sounds do the most work

Conversation runs on sounds that are barely words. “Yeah.” “Got it.” “Ooh.” A sharp intake of breath. Block names backchanneling explicitly as something Dessa learned from real conversation, and it is worth pausing on why that is hard for other systems: a transcript deletes these sounds almost by definition, so a model trained on text has never seen the signal that tells a speaker to keep going, speed up, or stop and check in. Dessa's demos lean on exactly these micro-behaviours — finishing a sentence when the speaker blanks on a title, gasping under its breath before answering, laughing at the moment something is funny rather than after a polite pause. You can hear the individual moments catalogued in what the official Dessa demos show. None of it is vocabulary. All of it is timing.

The volume trick, and what it proves about training

One demo detail deserves its own paragraph because Block flags it itself: ask Dessa to speak up and it gets louder — and, Block says, nobody taught it that. No instruction, no label, no text command in a training set, because there is no text in the training set at all. In recorded human conversation, “speak up” is reliably followed by louder speech, so a model that learns conversation from sound alone can pick up the behaviour the way a child does. The same logic explains the stress trick Block demonstrates with “I didn't say she stole my CUDA code” — stress a different word and the meaning changes, while the transcript stays identical. Timing, volume and stress are information that only exists in audio, which is the entire bet behind training on audio alone.

What we cannot verify yet

Honesty requires the closing caveat. Every timing behaviour described above comes from recordings Block selected and published. There is no public access, no independent latency test, no third-party session where a sceptic tried to talk over Dessa for ten minutes to see when it breaks. Block itself notes the model is occasionally awkward — in one recording it called a speaker “Ronaldo” — and awkwardness in conversation is usually a timing failure, which suggests the polished demos are a best case, not a guaranteed average. Dessa remains a research preview: early access for a small group of AI safety researchers first, a public waitlist, and no announced date for anything wider, as our availability guide tracks. When independent hands finally get hold of it, the first thing worth measuring is not what it knows. It is the clock: how fast it answers, how gracefully it yields, and how often it steps on the person it is talking to. That, more than any transcript, will tell us whether the timing in the demos is real.

Sources: Block's official Dessa page (block.xyz/dessa), re-read November 13, 2026; DessaAI Daily reporting linked above. DessaAI Daily is independent and unofficial, and is not affiliated with Block, Inc.

Frequently asked questions

← More Dessa AI articles Join the official waitlist ↗