Block's Dessa learned to talk from speech alone — and runs on a laptop

Every voice assistant you have used in the past three years hears you through a text layer: speech recognition, then a language model, then a voice synthesizer. Block's research lab says it built a model that skips all three steps — and that it fits on a laptop.

Block released Dessa on October 23, a full-duplex speech model trained from scratch on human conversation and nothing else — no text, no pretrained weights — that runs locally on a MacBook Pro M4 Max or better. The company published sample conversations and a waitlist, not a paper, a benchmark, or an API. R&D lead Ragavan Thurairatnam called it "a GPT-3 moment for speech" on X. That is a claim, not a result, and it deserves to be read as one.

What Block did publish is more interesting than most voice-model launches, because the design bet is unusual. Instead of bolting emotion onto a text pipeline with style tags — the approach ElevenLabs v4 and Fish Audio's Drama 3 take — Dessa never sees a word written down. It learned prosody, turn-taking and backchanneling from raw audio, which is why the demos have it laughing mid-sentence, dropping a quiet "oh my god" under its breath, or getting louder when the person on the other end asks it to speak up.

What it actually is

Audio in, audio out, full-duplex: it listens while it talks, so it can cut in and interject the way people do rather than waiting politely for a silence. Block says inference runs faster than real time, which is the difference between a conversation and a walkie-talkie. The whole thing was trained on a single NVIDIA GB200 NVL72 rack, according to Block's release post — small by frontier-lab standards, and a deliberate part of the pitch.

The training argument is the part worth taking seriously. Text is cheap and repetitive; speech is not. Say a sentence twice and the waveform still differs, and a single spoken word takes thousands of times more bytes than the same word typed. Speech also carries accents, pauses, breath, sarcasm and hesitation, and every one of those changes what a person means. Block's framing is that the industry has been throwing that signal away at the first step of the pipeline, and that a model trained directly on sound can keep it.

Block also claims Dessa learned language from "hundreds of times fewer words" than a comparable text model, since audio carries far more per token. There is no way to check that from outside: Block has not disclosed parameter count, training data volume, or hours.

The ceiling nobody is advertising

Every fact Dessa knows came from conversation, which means it knows a lot about how people talk and comparatively little about anything else. It cannot search, call a tool, or read a document. That constraint is visible in Block's own demo lineup: the most impressive clip is Dessa talking to GPT-Live-1, OpenAI's voice model, with a disclaimer noting OpenAI is not affiliated and has not reviewed Dessa. Two models, one conversation — the speech-native one providing the human sound, the other presumably providing the knowledge.

So the honest description of Dessa today is a voice, not an assistant. That is a narrower product than "the most authentic full-duplex speech model ever built," and it is also a more defensible one. A model that sounds like a person is useful in a call centre, a drive-through, a language tutor, or a support line even if it cannot answer a question about your account without help.

The commercial logic is not hard to find. Block owns Square, and Square already sells AI voice ordering to restaurants, TechCrunch notes. A voice model that runs on-device, with no per-minute API bill and no audio leaving the merchant's hardware, changes the unit economics of that business. Contrast it with OpenAI's hosted full-duplex model at five cents a minute — OpenAI puts its full-duplex voice model in the API at $0.05 a minute — where at a few thousand calls a day, that meter adds up. Local inference trades the meter for a machine.

The demo problem

Dessa is not available to use. There is a waitlist and an application form aimed at AI safety researchers, who will get early access in small numbers. No weights, no code, no licence, no latency numbers, no accuracy figures, no safety report. Every clip on the site is a curated recording made by the team, and the person in each recording is a Block employee or a colleague.

That is not disqualifying — frontier labs routinely show demos before papers — but it means the adjectives are doing work that evidence has not yet done. The field has a long record of "you cannot tell it from a human" launches that fell apart under blind listening tests, background noise, or a five-minute conversation about something the model did not know. The interesting engineering claim here, speech-only training, is exactly the kind of claim that needs an ablation: how much of the naturalness comes from the data, and how much from the fact that Block chose the good takes.

The second unproven claim is the hardware one. "Runs on a MacBook Pro M4 Max or better" is a strong statement about model size and memory, and it sits awkwardly beside "frontier speech model" — those two facts together imply a very compressed architecture. Block has not said how much memory it takes, how many milliseconds of latency it adds, or what happens to quality when the laptop is also running a browser. Anyone who has tried to run a local speech model on a fanless machine will want those numbers before believing the demo is reproducible.

The risk Block is already naming

A voice that carries emotion and a voice that fakes it are the same technology, and the difference is intent. Commentators on X went straight to the obvious use case — scam calls that no longer sound like scam calls — and CoinDesk framed its coverage around exactly that split between awe and fear. Block's answer is the researcher access programme, which is a tacit acknowledgement that it has shipped something it cannot fully characterise yet.

There is also a subtler issue. Dessa reads emotional state from voice and responds to it. That is a feature in a tutor or a support agent and a manipulation surface everywhere else, and no regulator has written rules for a model that adapts to your mood in real time. The company has said nothing about watermarking or provenance for audio it generates, which is the first question any deployment in payments or customer service will face.

What to watch

Three things would settle this. First, whether the safety access turns into a published evaluation with numbers rather than testimonials. Second, whether Square ships Dessa into voice ordering, which would make it the first speech-only model in production at scale and give outsiders a real workload to measure. Third, whether Block releases weights or an API — a local model nobody can download is a demo, and the entire pitch rests on the model being small enough to hand out.

Until one of those happens, the correct read is that Block has made the most interesting argument in voice AI this month and has not yet shown its work. The argument is worth taking seriously: the text bottleneck is real, and a model that never learned to read is a genuinely different object from one that did.

Would you trust a voice on the phone more or less if you knew it was trained only on speech? Write to us via our contact page.