Block's Dessa Was Trained on Speech Alone — Here's What the Demos Don't Prove

Block published more than a dozen conversation recordings and not one benchmark number. That ratio is the most informative thing about Dessa, the full-duplex speech model Jack Dorsey's company announced on 23 October 2026. Dessa is trained from scratch on human conversation, with no text and no pretrained models, and it does something genuinely unusual: it listens and talks at the same time, audio in and audio out, with no transcription step in the middle. It is also, as of this writing, a research preview behind a waitlist with no parameter count, no API, no price, no weights, and no published evaluation. The demos are the entire public evidence base.

That does not make the announcement empty. A speech model that learns language without a text corpus attacks an assumption most of the industry has stopped questioning — that text is where the knowledge lives and speech is a thin interface on top. But the gap between "the demos are startling" and "we know what this model can do" is wide, and Block's own page is more candid about that gap than most launch posts. The sentence that does the most work is four words long: "as far as we know."

What Block actually put on the page

Strip out the adjectives and the confirmed record is short. Everything below is from Block's own announcement or the company's public statements; nothing here has been independently verified.

• Release — announced 23 October 2026 as a research preview, not a generally available product.

• Form — end-to-end speech-to-speech: audio in, audio out, full-duplex, no text intermediate.

• Training — from scratch on human conversation only. Block says no text data and no pretrained models were used.

• Compute — training ran on a single NVIDIA GB200 NVL72 rack, according to the company.

• Local inference — Block claims it runs faster than real time on a MacBook Pro M4 Max "or better."

• Access — early access for "a small group of AI safety researchers" who apply with an institutional email, plus a public waitlist. The page says launch links are coming soon.

• Missing — parameter count, context length, dataset size or hours of audio, training duration, latency in milliseconds, benchmark scores, licence, price, weights, and any API documentation.

The team is a small research group inside Block led by Ragavan Thurairatnam, built around former members of the Toronto startup Dessa that Block acquired in 2020, and it reports directly to Dorsey. That lineage matters for reading the announcement: this is a lab with an eight-year history in generative speech, not a payments company bolting a demo onto a product roadmap. It also explains why the release looks like a research artifact rather than a product launch — because it is one.

Full-duplex is the part worth understanding

Most voice assistants work like a walkie-talkie. The system waits for you to stop, decides you have finished, transcribes, thinks, synthesises a reply, and plays it. Every one of those boundaries is audible. You learn to wait for the beep, and the model learns nothing from the half-second where you were about to speak but changed your mind.

Full-duplex is the phone-call model. The model is listening while it speaks, so it can be interrupted mid-sentence, can interject "yeah" or "got it" while you talk, and can adjust its timing to yours. Block's demos lean hard on this: Dessa laughs in the middle of its own sentence, raises its volume when asked, finishes a film title the speaker has forgotten, and gasps at an anniversary reveal. The Verge, which listened to the same clips, described the conversations as "a little awkward and uncanny" and flagged one where Dessa calls its conversation partner "Ronaldo."

Both reactions can be true. Full-duplex is an interaction property, not an intelligence property. A model can have flawless turn-taking and nothing interesting to say, or a brilliant answer delivered with the timing of a fax machine. What the demos establish is that Block has built something that behaves like a participant. What they cannot establish is how often it behaves that way when nobody is selecting the clip.

What training on speech alone buys, and what it costs

The case for speech-only training is that text throws away almost everything a conversation carries. A transcript of "fine" is one word. The audio is a sigh, a pause, a pitch drop, and a decision about whether to ask again. Block's stated safety motive follows from that: a voice system that cannot hear distress is deaf to exactly the signal that matters in a support call or a crisis line.

The cost is equally structural. Large language models get their factual coverage, their code, their structured output, and their tool calling from text. A model trained only on speech has no native interface for a PDF, a JSON schema, or an API call. Block concedes the open questions on its own page: how efficiently a speech-only model learns facts and languages compared with a text model, and whether in-context learning survives without text pretraining. Those are not footnotes. They are the difference between a compelling conversational agent and a general-purpose assistant.

There is a defensible position in the middle: a speech-native model does not need to read your spreadsheet if it is paired with something that does. But that is an architecture argument Block has not made yet, and the current preview exposes no interface through which such a pairing could happen.

The evidence base is a curated demo reel

Block says the clips are live recordings, that nobody steers the model, and that subtitles were added afterwards. Those are good practices and they rule out the crudest form of demo fraud. They do not rule out selection. A team that records many conversations and publishes the best of them has told you that good conversations exist, which is not the same claim as "this model is reliable."

What an independent evaluation would need is specific, and worth writing down now so the next announcement can be measured against it:

• A held-out conversation set with published prompts, not selected highlights.

• Response latency measured end to end, including the time to first audible reaction, not just "faster than real time."

• Interruption recovery: how often the model loses the thread after being cut off, and how quickly it re-enters.

• Intelligibility under noise, which is where speech systems usually fail first.

• A stated compute and data budget, so the "hundreds of times fewer words" claim can be checked against something.

• Some measure of factual retention, since that is the explicit trade-off of dropping text.

None of these exist yet. That is normal for a preview and abnormal for a launch, which is why the correct label for Dessa today is "research result with an unusually good demo page."

The hardware claim is the quiet one

Two hardware numbers sit on the page and neither has been checked. The first is that training used one GB200 NVL72 — a single rack holding 72 Blackwell GPUs. If true, it is a striking data-efficiency claim: a model that holds a conversation, trained on one rack, in a field where frontier text models are trained on tens of thousands of accelerators. It is also the kind of claim that depends entirely on scale: one rack for how long, on how many hours of audio, with how many parameters?

The second is that inference runs faster than real time on a MacBook Pro M4 Max. A top-end M4 Max has 128 GB of unified memory, so a model that fits comfortably there is not enormous by modern standards. That is an inference, not a disclosure. Block has not published the size, the quantisation, the memory footprint, or the tokens-per-second figure, so the laptop claim cannot be reproduced, only repeated.

This is the place where a technical reader should feel the absence most sharply. Everything about the model's practicality — cost per conversation, whether it can run on a phone, whether an API could ever be cheap — follows from a number Block has chosen not to publish.

Why access is gated, and why that matters for the eval

Block's stated reasons for building Dessa are curiosity, a better understanding of human–AI interaction, and safety. The gate is consistent with the third: a voice that reads emotion and converses convincingly is a ready-made impersonation tool, and the fastest way to avoid a bad headline is to keep it in a small room.

The trade-off is that gated access and independent evaluation pull in opposite directions. A small group of safety researchers may well produce the most useful red-team results, but those results will arrive on Block's timeline and, in most cases, under Block's review. Until an outside group publishes numbers, the public record consists of the vendor's clips plus journalists' impressions of the vendor's clips. That is a thin basis for the phrase "surpasses anything else in emotional expression," which is a claim about a comparison nobody has run.

What would turn this into a product

Four things, in rough order of importance. A price or a weight file, because either one tells you who can use it. An API or an SDK, because a model you cannot call is a paper. A benchmark or an independent eval, because "most authentic" is not a metric. And a latency figure in milliseconds, because the entire value of full-duplex is timing and timing is the one thing a video clip cannot convey.

Any one of those would change how this article should be read. All four would make it a launch.

If you need a voice agent before then

The practical stack for a production voice agent is still speech recognition, then a language model, then speech synthesis. Each hop adds latency and each vendor is a separate contract. OrcaRouter is the middle layer: one API in front of 200+ models, with 0% markup — provider list price passed through — automatic failover across providers, a routing DSL that can compose several models into one call, and model fusion when you want a panel of models answering together. That does not make a walkie-talkie into a phone call; it removes one integration from the chain while the duplex models mature.

To be explicit about what we do not have: Dessa is not in the OrcaRouter catalogue, and nothing in this article should be read as an offer to serve it. When Block ships an API, the routing layer is where a model like this would be worth comparing on price per minute of conversation rather than per token.

What to watch

The claim worth taking seriously is not "first." It is that language might be learnable from speech at hundreds of times less data, and that the resulting model might understand a conversation in a way a transcript never could. If Block publishes a paper with the data budget, the parameter count, and a held-out evaluation, that claim becomes testable. If the next announcement is another set of clips, the interesting question will still be open — and the waitlist will still be the only door.