Block Dessa: Jack Dorsey's Voice AI Trained Only on Speech

Block Dessa is the voice AI everyone was clipping on Friday. Jack Dorsey's Block unveiled Dessa on October 23, a full-duplex speech model it says was trained "from scratch on nothing but human conversation," with no text and no pretrained models. It listens and talks at the same time, laughs mid-sentence, and runs locally on a MacBook Pro M4 Max or better, per Block.

What actually happened

Block published a Dessa landing page packed with live recordings of people talking to the model, with subtitles added afterward. Block says nobody steers it: Dessa picks what to say, when and how, including its own voice and personality.

The pitch is that most voice assistants turn your speech into text, think in text, then turn the answer back into audio. Block argues that pipeline throws away tone, pauses, sarcasm and emotion. Dessa is "audio in, audio out," a true speech-to-speech model where everything it knows came from recorded speech.

Because it is full-duplex, it can interrupt, wait, or drop short backchannels like "yeah" and "got it" while you are still talking. Block says inference runs faster than real time, so there is no awkward lag.

Who built it

The page says Dessa comes from a "very small" AI research lab inside Block, founded by former members of Dessa, the Toronto deep learning startup. The team reports directly to Dorsey, and the project lead is Ragavan Thurairatnam.

Block gives three reasons for building it: curiosity, better understanding between humans and AI, and safety. Its safety argument is that today's AI is "nearly deaf" to emotion in our voices, so a voice agent might not notice someone in serious pain. Block frames the Block Dessa project as a step toward fixing that.

The demos, honestly

The Verge listened to the clips and found them impressive but odd at times. Beyond the "Ronaldo" mix-up, one chat about Christopher Nolan films drifted into a tangent about Bryan Singer's Superman movies. Another clip shows Dessa reacting to a surprise trip to Paris with what Block describes as "teary gratitude."

Block also posted a conversation between Dessa and OpenAI's GPT-Live-1 voice model. Block notes the other developer is not affiliated and did not review the demo. The Verge says Dessa comes across as the more natural speaker, but that is a single cherry-picked clip, not a benchmark.

Demo reels are marketing. Until outsiders can test a model on their own prompts, "most human-like" is a claim, not a result.

Why Block Dessa matters for AI agents

Voice is becoming a big front for agents, from customer support to in-game characters. A model that hears hesitation, frustration or sarcasm could, in theory, make agents safer and less annoying. A model that sounds this human also raises obvious risks, from scams to people forgetting they are talking to software.

Block seems aware of that. It is opening early access to safety researchers first, and it says a model that learned to talk from sound alone is "new territory." For more on how fast voice AI is moving, see our Mistral Voxtral 3 explainer and our HeyGen Voice TTS piece.

What to do as a reader

If you want to try it, the only official path right now is Block's own page. Block has a request-access form for safety researchers that asks for an institutional email, and says more recordings will be shared on X soon. Ignore anyone selling "Dessa access," a Dessa app, or a Dessa token. Block has not announced any of those.

If you build voice products, watch for the things that matter: published evaluations, a technical report, license terms and hardware needs beyond a top-end MacBook. Until then, Block Dessa is a fascinating preview, not a tool you can ship with. And no, a viral AI demo is not a reason to ape into the first ticker that copies its name. Not investment advice.

Sources: Block's Dessa page and The Verge.