Prince Mario-Max Schaumburg-Lippe: Block's Dessa Learned to Talk From Voices Alone

Most voice assistants work like a relay race. One program turns your speech into text. A language model reads the text and writes a reply. A third program reads that reply out loud. It works, after a fashion. It also throws away almost everything that makes a voice a voice – the pause before bad news, the laugh that arrives half a second early, the sigh that says more than the sentence did.

Block, the payments company co-founded by Jack Dorsey, unveiled a different bet on October 23, 2026. The model is called Dessa. According to the lab behind it, Dessa was trained from scratch on nothing but recorded human conversation. No text. No pretrained models underneath. Audio goes in, audio comes out, and at no point does anything get flattened into words on a page.

Listening and talking at the same time

Dessa is full-duplex. That is the technical term, and it describes something you do every day without thinking about it – listening while you talk. The model can laugh mid-sentence. It can cut in when a thought lands. It can backchannel with a quick “yeah” or “got it” while the other person keeps going. Anyone who has waited through the stiff, walkie-talkie rhythm of a normal voice assistant knows how much of conversation lives in exactly those overlaps.

The lab says Dessa natively hears and expresses emotion, sarcasm, laughter and the small cracks in a voice that signal how someone actually feels. In one published recording, it finishes a movie title the speaker cannot quite recall. In another, it reacts to a surprise anniversary trip to Paris with what sounds like genuine, teary delight. A third recording has it in conversation with another AI voice model, holding its own end of the exchange.

Those clips are the evidence on offer, and the lab is open about that. Each published recording, Block says, is a live conversation with the model, with subtitles added afterward. Nobody steered the replies. Still, recordings chosen by the maker are recordings chosen by the maker. Treat them as a promising demonstration, not a scorecard.

Trained on one rack, running on a laptop

Two numbers give the project its shape. Block says training ran on a single NVIDIA GB200 NVL72 system – a serious piece of hardware, but a single rack rather than a sprawling data center build-out. And the finished model, the lab says, runs locally on a MacBook Pro with an M4 Max chip or better, with inference faster than real time. That second point matters more than it might sound. A voice model that lives on your own machine does not need to send your conversation anywhere. It can simply sit there and talk.

The lab also says Dessa learned language from hundreds of times fewer words than large language models typically need. No dataset size has been published, so that claim stays in the lab’s column for now. The same goes for the bigger framing. Dorsey called Dessa “the most authentic full-duplex speech model ever built” in a post viewed more than 6.7 million times. Authentic is a word the coming months will have to earn.

Readers following this week’s open-model news will spot a friendly contrast in Kimi K3.1 Pro arriving with open weights promised for November 4. That model is built for text and code at enormous scale. Dessa points the other way – small, local, and built entirely out of sound. Both bets say something about where builders think the value sits now: less in raw size, more in fit for the job.

A small lab with an eight-year head start

Dessa comes from a very small research lab inside Block, founded by former members of Dessa, the Toronto deep learning startup, and led by Ragavan Thurairatnam. The team reports directly to Dorsey and says it has spent more than eight years building generative speech models. Eight years is a long apprenticeship in a field that reinvents itself every eighteen months. It shows in the priorities. The team lists three reasons for the work: curiosity, better understanding between people and AI, and safety – a voice system that can hear distress in a caller’s voice, they argue, is a safer system than one reading a transcript.

That safety angle explains the rollout. There is no public access yet. Early access goes first to a small group of AI safety researchers, with a public waitlist open behind them. Block has published no parameter count, no API, no pricing, no weights and no benchmark results. Independent coverage found the demos mostly natural, with the odd awkward moment – in one clip the model mistakes the speaker for someone named Ronaldo. Honestly? A small stumble like that almost helps the case. Real conversation is full of them.

What to watch next

Three things will decide whether Dessa becomes a product or stays a beautiful research preview. First, outside evaluation – safety researchers with hands on the model will say things a launch page cannot. Second, consistency over long conversations, where small timing errors pile up. Third, whether local, laptop-grade speech starts showing up in ordinary tools, the way Waymo starting driverless rides at Las Vegas airport turned a long-running experiment into a service you can simply book.

For now, the idea itself is the news. A machine that learned to talk the way people learn to talk – by listening, for years, to other people talking. No books. No web pages. Just voices. If the recordings are even half representative, the next time a voice assistant laughs at the right moment, you may not notice at all. And that, quietly, is the point.

Also published on: