Block's Dessa is a speech-to-speech AI trained from scratch on human conversation

Block enters the AI model race with Dessa

Block, the payments company associated with Jack Dorsey, has announced Dessa, a conversational AI model that the company says was trained from scratch using nothing but human conversation. According to The Verge, which reported the announcement, Block published a batch of recorded chats to show the model in action — exchanges that come across as mostly natural, with moments that are noticeably awkward or uncanny.

What the demos show

In one conversation shared by Dessa lead Ragavan Thurairatnam and described by The Verge, the model sustains a largely fluent exchange with a speaker, though it initially misidentifies the person as someone named Ronaldo. Other examples on the Dessa website include a discussion of Christopher Nolan's films that wanders into an unrequested detour about Bryan Singer's Superman movies, a moment where a user describes a surprise trip to Paris and the model reacts with what the site characterizes as teary gratitude, and a sit-down between Dessa and OpenAI's GPT-Live-1 voice model in which, as The Verge observes, Dessa comes across as the more natural conversationalist.

A model built on speech alone

The core technical claim is that Dessa is, in the team's words as quoted by The Verge, a "true speech-to-speech model" — everything it knows came from human speech, not from text corpora converted into audio. That implies no intermediate step where spoken language is flattened into plain text before the model responds. The team also says the model can run locally on a MacBook Pro M4 Max or better hardware, which would put a full conversational model on a consumer laptop without depending on cloud inference.

Why Block says it built Dessa

The Dessa team gives three motivations: curiosity, better mutual understanding between humans and AI, and safety. Expanding on them, as reported by The Verge, the team points out that babies learn to talk without ever seeing a written word, and says they wanted to find out whether a model could do the same — it can, they claim, but it is very hard. They also argue that speech is the most natural interface to any intelligence, and that today's voice AI squanders it by reducing what people say to text and failing to keep up with human timing.

The safety argument is the sharpest. As AI systems gain power and independence, the team writes, they remain almost deaf to the emotion carried in human voices. Their example imagines a person in serious pain while a voice assistant carries on, completely unaware — a gap the team presents Dessa as a step toward closing. Early access will be limited to a small group of AI safety researchers, according to The Verge.

Why it matters

Most voice assistants in production today are pipelines: audio is transcribed, a text model formulates a reply, and a synthesizer reads it back. That design discards prosody, timing and emotional cues — precisely the information a speech-native model like Dessa is built to retain. If the approach holds up beyond vendor-curated demos, it points toward voice interfaces that can register hesitation, distress or excitement instead of treating every utterance as flat text.

It is also a notable competitive signal. Block is primarily a fintech company, and training an in-house model from scratch places it in a race that has so far been dominated by dedicated AI labs. The claim that the model runs on a laptop, meanwhile, fits a broader shift toward capable models that do not require cloud data centers. The caveats are real — the published conversations were selected by the team, the Ronaldo mix-up reveals rough edges, and there is no word yet on general availability — but as a research direction, building AI that learns language the way infants do, through speech alone, is worth watching.