Dessa Learned to Talk Without Ever Reading a Word
- What is actually different
- The part that matters more than it sounds
- Where I would expect it first
- Where this goes
- What is not here
- Sources
A model at Block learned to talk without ever reading a word.
Not fine-tuned on transcripts. Not trained on text and wrapped in a voice. Trained from scratch on human speech alone: audio in, audio out, one model, nothing in between.
Block calls it Dessa. The team behind it came out of the Toronto deep-learning startup Dessa, has been building speech models for over eight years, and reports directly to Jack Dorsey. Their claim is a first: a system you can hold a live conversation with that learned language only by listening.
What is actually different
Every voice assistant you have used does the same thing. It hears you, converts your voice into text, hands the text to a language model, and speaks the answer back out. That pipeline is why talking to machines feels like filling in a form.
Everything that is not words is thrown away at the first step: the pause before a hard sentence, the sarcasm, the sigh, the way a voice cracks. The words survive. The meaning does not.
Dessa skips the text entirely. In the published recordings it laughs mid-sentence, the way people do, rather than after you finish. Asked to speak up, it gets louder. Nobody taught it either behaviour, as the makers put it, and I have no way to verify their account of how it emerged.
It is also full-duplex. It listens and talks at the same time, cuts in, and makes the small noises that signal attention, the “yeah” and the “got it”, instead of waiting politely for silence. Inference runs faster than real time, and it runs locally on a MacBook Pro with an M4 Max or better. The maker says the whole model was trained on a single NVIDIA GB200 NVL72.
The part that matters more than it sounds
A system that never sees text gets to ask a different question. Not how smart it is, but how efficiently language and facts can be learned from sound alone, and how fast a new language arrives when the only teacher is speech. Block states that Dessa learned from hundreds of times fewer words than a language model needs, a claim from the maker rather than a measurement.
It also had to learn more than language. Accent, breathing, inflection, timing, when to interrupt. Speech never repeats: say a sentence twice the same way and the data still differs. The makers call that part brutal, and it is the reason the approach is uncommon rather than merely late.
Where I would expect it first
This section is my own reading, not a claim from Block.
Not in a chatbot. In the places where the voice carries the information and the words do not. Emergency calls, where someone may be unable to say plainly what is wrong but will always sound like it. Crisis lines, where the difference between a flat “I’m fine” and one that isn’t lives in the delivery. Customer service at the moment frustration turns into churn, which Block argues the current generation of voice AI cannot perceive at all.
And a fourth one that matters here: agents. A voice layer that carries tone and interruption would end the artificial rhythm of speech, think in text, speak again. The conversation would finally be a conversation instead of two dictations with a delay between them.
Where this goes
The interesting line is not better voices. It is that the text layer, the thing everyone assumed was necessary, turns out to be a lossy middleman. If that holds, the next generation of speech systems is not built on top of language models. It is built beside them, on sound, with its own scaling behaviour and its own failure modes that we have no names for yet.
Running locally, on a laptop, faster than real time is the second part. A model that only ever touches audio produces no transcript of you in the first place. For anything that listens at all, in a doctor’s room, a help desk or a home, that removes the most obvious record. It does not remove every kind of telemetry, and the page does not claim it does.
What is not here
Nothing is downloadable. The page says launch links are coming soon. There are no weights, no API, no benchmark anyone outside has run. The safety team is explicitly looking for outside researchers to find what they missed, which reads like the honest position for a model trained in a category that did not exist before.
The recordings are persuasive and I cannot check them. The hardware claim is a laptop I do not own, and the adjective in the headline, the most authentic full-duplex speech model ever built, is the maker’s and not a finding.
What is defensible tonight is the argument rather than the product. A system that hears speech through text has already decided to ignore most of what you said.
Write a comment