Block's Dessa is a voice AI trained only on human speech, and it laughs and interrupts like a person
Block's AI research lab unveiled Dessa on Oct. 23, 2026, a full-duplex speech-to-speech model it says was trained from scratch on nothing but recorded human conversation, with no text and no pretrained models, released as a research preview with demo recordings, a public waitlist and early access for a small group of AI safety researchers.
Almost every voice assistant today is a text model wearing a voice costume: speech in, transcript, text answer, speech out. Dessa skips the text entirely, which is why it can laugh mid-sentence and pick up on sarcasm. That's a genuinely different bet, and the fact it runs on a laptop makes it more interesting, and more worrying. A voice that sounds this human, running locally, is a scammer's dream, which may be exactly why Block is handing it to safety researchers first. The skeptic's note: no benchmarks, no model size, curated demos, and a "GPT3 moment" claim from the team that built it.
Block says Dessa is a speech-to-speech model trained from scratch only on human speech: no text, and no pretrained models. Speech-to-speech means a person talking goes in as audio and speech comes back out, with no written transcript in the middle. Trained from scratch means it did not start from a model that had already learned language off text. A pretrained model is one of those head starts. Those lines are Block's, on its Dessa page.
It is full-duplex, which means it listens and talks at the same time, the way a phone call works, instead of waiting its turn. Block says it knows when to cut in and when to backchannel, the little sounds that show a person is still listening, such as "yeah!" and "got it!". It hears and produces emotion, sarcasm, and laughter as part of the speech, not as a label added later. Block says inference, the step where the model answers, runs faster than real time. Those lines are Block's.
It runs locally on a MacBook Pro with an M4 Max chip or better. Locally means on that laptop, not on a server in a data center. The M4 Max is a high-end Apple chip. TechCrunch reports that Block trained Dessa on a single NVIDIA GPU rack, a cabinet of graphics chips used to train a model. CoinDesk reports that same training run as a single Nvidia GB200 NVL72, NVIDIA's name for one of those racks. Those hardware lines are as the two outlets report them from Block.
Lab lead Ragavan Thurairatnam called it "a GPT3 moment for speech" on X. Jack Dorsey, Block's co-founder and chief executive, wrote that talking with it "feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence." Those quotations are theirs.
There is no public access yet. Dessa is a research preview, which means a look at the work, not a product you can switch on. Block has a public waitlist, and early access is going to a small group of AI safety researchers. Block has not disclosed the model's size, any benchmark scores, or a release date. A benchmark is a shared test other people can rerun. The demos are recordings Block made, not tests run outside the company. Those limits are what Block has put on the page.
CoinDesk described the reaction as split, between awe at how human the voice sounds and worry that it could make phone scams harder to spot. That split is CoinDesk's read of the response.
The picture is Steve Jurvetson's photograph of Block co-founder and chief executive Jack Dorsey, seated on stage in a black T-shirt against a purple curtain. The credit is Steve Jurvetson, via Wikimedia Commons, under CC BY 2.0.
In plain terms, Block said on Oct. 23 that Dessa learned to talk by listening only to people, with no text in the loop, and that it can laugh, get sarcastic, and interrupt. It runs on a high-end MacBook. You cannot use it yet. The public step is a waitlist, and a small group of safety researchers gets it first. Block has not published a model size, a score, or a date.