Block's Dessa voice AI listens and talks at once, and runs on a laptop

What Block announced

Jack Dorsey, the former Twitter CEO, introduced a speech model called Dessa from his company Block. As indianexpress.com reported, Dorsey announced it on X, calling voice the most natural way to deal with an AI system and describing Dessa as the most authentic full-duplex speech model built so far.

Dorsey also stressed its size. In his post he said the model is "small enough to run on your MacBook." Block says it runs faster than real time on a MacBook Pro M4 Max or better, with no noticeable delay.

Access is tightly limited. Dessa is a research preview behind a waitlist. There is no price, no API (a way for other software to plug into it), no published weights (the learned numbers that make up a model) and no parameter count, which is a common measure of model size. No formal evaluation has been published either.

Listening and talking at the same time

"Full-duplex" means both sides can speak at once, like a phone call rather than a walkie-talkie. Most voice assistants today wait for you to finish, turn your speech into text, work out a reply and then read it aloud. Dessa, by Block's account, skips that. Audio goes in and audio comes out, with no transcription layer between.

Block's argument is that the text step throws away a lot. Tone, hesitation and mood do not survive being written down. The Dessa website says the most powerful AI models cannot hear how a user feels, which the company says makes for a worse experience and can even be dangerous.

The demo recording on the site shows what Block means. Dessa laughs in the middle of a sentence when something is funny, rather than waiting and laughing on cue. It fills in when the speaker loses their words and gasps like a person would. The company also posted a demo of Dessa talking with ChatGPT Live.

Trained on sound alone

The most unusual claim is how Dessa was built. Block says it was trained from scratch on real human conversations, with no text and no pretrained models. The website says it learned to hold a live conversation just by listening, and describes it as the first realistic live-conversation AI trained only on human speech, as far as the company knows.

That is harder than it sounds, according to Block. Text is the same no matter who writes it, but speech varies even when a person says the same sentence twice. A single spoken word also takes thousands of times more bytes of data than the same word typed. On top of language, the model had to pick up accents, pauses, breathing, humour, sarcasm, inflection and emotion. It also had to learn when to stop talking.

Block says it found a way to train the model on only one NVIDIA GB200 NVL72, a large piece of AI computing hardware. That is a company claim, and no outside party has checked it.

Who is behind it, and what is unproven

Dessa comes from a small AI research lab inside Block. The lab was founded by former members of Dessa, a deep learning startup from Toronto, who have spent years building deep learning and generative speech models.

Block also frames the model as a research tool. The website raises open questions it hopes Dessa can help answer. How efficiently does a model learn language and facts from speech compared with text? How quickly can it learn new languages? What kinds of in-context learning, meaning picking up a task from examples in the conversation, can it do?

For now, the only evidence is the public demos. Demos show a best case, so the claims of no delay, natural emotion and laptop-speed performance remain unconfirmed until independent testers get access.

What to watch from here

The most useful signals will be practical ones. Does Block open the waitlist, publish test results, name a price or release the model for others to build on? Those steps would show whether Dessa is a research showcase or a product.

It also matters that this comes from a company better known for payments than for AI research. It suggests human-sounding voice is no longer limited to the biggest AI labs.