Learning to talk by listening alone: Block's Dessa brings speech models back to sound itself

The short version: this is not just another voice assistant

Block's Dessa, unveiled on October 23, 2026, needs one correction up front: it is not a nicer voice bolted onto an existing chatbot, nor another tool that turns speech into text and reads the answer back. By Block's own account, Dessa is a true speech-to-speech model: audio in, audio out, with no text in the middle. It listens while it talks, so it can interrupt, cut in, and backchannel with short words like "yeah" and "got it" to show it is following along, and it can laugh mid-sentence when something is funny. Block calls it "the most authentic full-duplex speech model ever built." That phrase is vendor language, but the training method behind it genuinely deserves separate discussion: Dessa used no text corpus and no pre-existing large language model. It learned to talk only from recorded human conversation.

The distinction sounds technical, but it affects the most basic user experience of voice products. Most voice assistants today follow a cascade: speech is recognised as text, the text goes to a language model, and the model's text answer is synthesised as speech. Every step loses something. Tone, pauses, hesitation, laughter, sighs, and sarcastic intonation either do not exist in text or survive only with extra annotation. Block's argument is that this "compress it into text first" approach is exactly why voice assistants sound polite and correct yet always feel as if they are reading aloud. Dessa does the opposite: it treats the things that would normally be thrown away as the learning material itself. What it must learn is not only what was said, but how it was said.

The qualification matters just as much. Dessa currently has no public access, and Block has published no parameter count, API, price, weights, or benchmark results. What the public can see is a set of live recordings released by Block plus press coverage. So this article discusses a research preview and a claim, not a product you can download and integrate today. Its significance is not "usable now" but the question it makes testable: how far can a model that never touches text get at learning to talk? Block's first evidence at least makes that question concrete rather than hypothetical.

What it actually is: full-duplex, speech-to-speech, listening only

Three terms need unpacking. First, "speech-to-speech": Dessa's input and output are both audio, with no transcription station in the middle handing text to a language model. Block's page puts it bluntly: "Sounds! No text, no pretrained models." The implication is that everything the model knows came from conversations it heard, not from web pages, books, or code it read. Second, "full-duplex": ordinary voice assistants are mostly half-duplex — you finish, then they answer, and an interruption mid-answer is often missed or deferred. Dessa is designed to work like a human conversation in both directions at once: it keeps listening while it speaks, so interruptions, overlaps, and backchannels are normal parts of dialogue rather than edge cases to handle specially. Third, "trained only on speech": Dessa was trained from scratch on human conversation alone. It is not a text model wrapped in a voice shell, and that distinction sets the premise for everything that follows.

Block also gives two specific claims about scale and efficiency, both of which must be labelled as vendor claims. First, training used a single NVIDIA GB200 NVL72 rack, not a cluster of hundreds or thousands of machines. Second, Dessa learned language from "hundreds of times fewer words" than large language models see. Block itself adds a fair caveat: language models also have to learn C++ and everything else, so a word-count comparison is not apples to apples. Even discounted, the two numbers point at the same judgement: speech signals differ from text in information density and learning difficulty. A spoken word takes thousands of times more bytes than the same word in text, and the same person saying the same sentence twice never produces an identical waveform. The model must learn accent, rhythm, breathing, and emotion inside a signal that never repeats exactly. That is the "hard" part Block keeps stressing: babies learn to talk without ever seeing a written word, but getting a machine to do it with no textual crutch is something that had not really been done before.

On product form, Dessa is deliberately restrained. Block says it runs locally on a MacBook Pro M4 Max or better, with inference faster than real time, so answers do not need a cloud round trip. Early access goes only to a small group of AI safety researchers, to find risks the lab has not anticipated; everyone else joins a waitlist. That ordering is itself a signal: Block knows a voice model that sounds this human carries risk on the same side as its appeal. It chose to hand the model to people whose job is finding problems before opening it widely. That deserves credit, but it also means any judgement about "how well it really works" currently rests on recordings Block selected, not on independent evaluation.

What is actually worth watching in the demos

Block's demos are not polished advertisements but labelled live recordings of people talking to Dessa, with subtitles added afterwards. The lab stresses that nobody steers the model mid-conversation: Dessa decides what to say, when to say it, and in what voice and personality. A few details in those recordings matter more than the overall verdict of "sounds very human."

The first is the immediacy of emotion. In one clip the model says it is tired from work, and it sounds tired. In another it hears a surprise, gasps quietly, then answers with what sounds like tearful gratitude. In another it laughs in the middle of a sentence rather than waiting for the other person to finish and then laughing on cue. None of this is achieved by inserting a "[laugh]" tag into text. A cascaded system trying to reproduce it would need to recognise the emotion, encode it as a tag, and have the synthesiser perform it from the tag — a longer chain with more room for distortion. Dessa learns emotion as part of language rather than as decoration pasted on afterwards.

The second is conversational repair. In one demo the other person blanks on a movie title and Dessa finishes the sentence. In another the person asks it to speak up, and it genuinely gets louder, with nobody having taught it which parameter "louder" maps to. These behaviours look trivial in text chat, but in voice they are decisive: real conversation is full of small negotiations that are never stated explicitly and are completed from context. Whether a model can handle those micro-negotiations determines whether it sounds like an assistant or like a person. Block's claim is that Dessa can do this because it learned timing rules from real conversation — when to cut in, when to backchannel — not just word sequences.

The third is its conversation with another AI. Block released a recording of Dessa talking to OpenAI's GPT-Live-1, with an explicit note that GPT-Live-1's developer is not affiliated with Block and did not endorse Dessa. The point of that clip is not who won but what it reveals at the same time scale: interruptions, overlaps, and pauses are amplified when two systems talk to each other. A single-person demo can hide rehearsal; a two-AI conversation exposes whether rhythm control is genuinely running in real time or is a rehearsed monologue. Block's demos still have limits: selected clips naturally favour successes, and press coverage notes occasional awkward moments, such as mistaking the speaker for someone else. Those do not refute the work. The real problem is the absence of public benchmarks, independent evaluation, and a reproducible access route, so the public cannot tell what share of all conversations the successful clips represent. And Block itself concedes that Dessa's knowledge comes entirely from heard conversation, which naturally limits factual knowledge: it may be a great conversationalist without being good at lookup, arithmetic, or coding. Equating "talks well" with "knows everything" is the easiest misreading in this news cycle.

Why Block, and why now

The team is small. Block says it is a lab founded by former members of Dessa, the Toronto deep-learning startup, that it has built generative speech models for over eight years, and that it reports directly to Jack Dorsey, Block's co-founder and head. Its lead, Ragavan Thurairatnam, called Dessa "a GPT-3 moment for speech" in a public post. That analogy should be discounted: GPT-3 became a moment because it quickly gained verifiable, widespread uses, while Dessa has no comparable external validation yet. But the small-team-big-problem combination explains Block's intent: it is not chasing an existing product category but betting on the next layer of voice-interaction infrastructure.

Commercially, the logic is easy to see. Block's main business is Square and other services for small merchants, and voice already appears in its ordering flows. A model that runs on-device, does not need every utterance sent to the cloud, and can hear emotion could in theory cut cost, shorten latency, and work where connectivity is poor. That is theory only. Block has announced no merchant product plan and has not said Dessa will enter Square's product line. Treating a research preview as a commercial roadmap is another common misreading. The accurate statement is: Block made a technical claim first; commercial landing is a question for later.

The timing matters too. Dessa arrives as voice assistants shift from "understanding commands" to "sustaining conversation," with OpenAI's GPT-Live as the most frequent comparison. Block's approach contrasts with that mainstream route: the mainstream puts voice capabilities around a strong text model, while Dessa asks the reverse question — if a model never gets text at all, what can it learn? That contrast has value on its own, because it reopens a premise that is usually taken for granted: we have long assumed speech must first become text to be understood, but that assumption may be an engineering convenience rather than a cognitive necessity. Dessa cannot yet disprove it, but it makes it discussable again.

What it does not prove: four red lines

First, it does not prove that speech models do not need text. It proves that, in Block's chosen conversational settings, a model trained only on speech can sound very natural. Natural does not mean general. Text models are powerful partly because text carries humanity's accumulated structured knowledge — mathematics, code, law, science — which is hard to learn from everyday conversation recordings alone. Dessa's knowledge boundaries have not been systematically measured, and any description of it as a "next-generation general model" exceeds the evidence.

Second, it does not prove that small compute can reproduce it. One GB200 rack sounds much smaller than a frontier cluster, but it is still top-tier hardware, and Block has not published training data scale, duration, or method in detail. Whether other teams can reproduce it with similar resources remains unknown. Reading "one rack" as "anyone can train one at home" misreads the number.

Third, it does not prove that local inference equals privacy safety. Local inference reduces the need to upload audio to the cloud, but the model itself, recording storage, conversation logs, and merchant-system integration all reintroduce privacy questions. More importantly, a model that hears emotion and mimics tone carries new risks of its own: it lowers people's guard more easily and can be used to impersonate a real person. Block's decision to give safety researchers early access shows it recognises these risks need study first, not default absence.

Fourth, it does not prove that "most realistic" means "best." In many settings users do not need an assistant that sounds like a person; they need one that is fast, accurate, and predictable. A model that interrupts you, sighs, and expresses emotion may be worse than a calm, concise system with a written record in customer service, healthcare, or finance, where accuracy and traceability matter most. Realism is a product choice, not a universal virtue.

Three things actually useful for vibe builders

First, design voice interaction sound-first, not text-first. Many vibe projects design voice features by writing the text dialogue flow first and then handing the text to a synthesiser. Dessa warns that this systematically discards information: if your product genuinely depends on tone, pauses, and interruptions to convey meaning, you cannot design dialogue only at the text layer. You must design rhythm, interruption rules, and emotional boundaries at the audio layer first, and only then choose a technical approach.

Second, build disclosure and traceability in up front, not as an afterthought. One detail worth copying from Dessa's release is Block's explicit labelling of which recordings are AI-generated and which conversation partner is a third-party model. For a vibe project, that translates into three hard rules: users must be able to tell at a glance that they are talking to AI; critical actions must never rely on voice confirmation alone but need a text or button second confirmation; and key conclusions from voice conversations must leave a written record, not exist only as a recording. These rules are not extra burden; they are the evidence trail.

Third, do not use demos as selection criteria. Dessa has no public evaluation, and that is a general lesson: when choosing a voice solution, what matters is not how natural a demo recording sounds but its failure modes in your setting. Run a small comparison: the same set of real user questions through a cascaded solution and a direct speech solution, and record interruption success rate, false-interruption rate, emotion misreading rate, and task completion rate. Without those numbers, any judgement of "sounds more human" is aesthetic, not engineering.

How this differs from our earlier voice coverage

This site has previously covered Sesame's voice assistant, Microsoft's voice stack, and Simon Willison's voice vibe-coding session, plus a practical voice-interaction guide. Those pieces landed on product updates, a vendor stack, a personal record, and selection method respectively. This piece lands somewhere else: a training paradigm — whether a model can learn dialogue without ever touching text. Dessa is not a competitor report against Sesame or Microsoft products, nor a sequel to the voice guide. It puts the premise "speech must first become text" back on the table as evidence. If you have read our voice-interaction guide, treat this piece as its reverse footnote: the guide tells you how to do things right within the existing cascaded architecture, while this piece reminds you that the cascaded architecture itself may not be the only answer.

Conclusion: a question worth taking seriously, not a ready answer

The thing most worth remembering about Dessa is not the adjective "most authentic" but the question it asks: if a model can learn to talk just by listening, was our habit of compressing speech into text helping the model understand, or making decisions on its behalf? Block's evidence so far is not enough to answer that question, but it is enough to make it concrete: live recordings, a description of the training method, a claim of local operation, and a safety-researcher-first plan. The three verifications that matter next are whether independent evaluation can reproduce the naturalness in the demos, where the model's boundaries lie on knowledge tasks, and what risks safety researchers find that the lab did not anticipate. Until those results exist, the honest stance on Dessa is: take its question seriously, and treat its adjectives with caution.