Block Builds a Voice AI That Never Read a Word

Almost every voice assistant you have used works by turning your speech into text, deciding what to say back, and then turning text into sound. Block, the company Jack Dorsey runs, has built one that skips the words entirely. Its new model, Dessa, listens to audio and answers in audio, with no detour through written language in between. The company announced it on 23 October, and the coverage since, from CoinDesk to a ten-outlet roundup, has swung between admiration and alarm.

The training approach is the headline. Block says it built Dessa from scratch on a single Nvidia GB200 NVL72 system, using only recordings of people talking. No text datasets, no borrowed pretrained model. The idea is that a model raised on raw speech learns the things text throws away, which is the timing, the overlap, the small noises and hesitations that make a conversation sound like two people rather than two monologues. Dessa is full-duplex, meaning it can listen and speak at once, so it can be interrupted and interrupt back.

The second claim is about size. Block says the model is compact enough to run locally on a MacBook Pro with an M4 Max chip or better, which would put a fluent conversational voice on a laptop with no round trip to a data centre. That is the sort of detail worth holding lightly for now. As one explainer pointed out, Block has not published latency figures, model size, or benchmarks, and no outside group has reproduced the results. For the moment Dessa is a research preview with a waitlist, and a small group of AI safety researchers has early access.

The part that makes people nervous

A voice this natural, running cheaply on ordinary hardware, is exactly what the people who track phone scams have been dreading. A full-duplex model that can hold its end of a live call, hesitations and all, erodes the last easy tell that you are talking to a machine. CoinDesk's headline caught the mood, describing a voice that "sounds eerily human" and the fraud worries that came with it.

It is a real tension, and I do not think Block has resolved it by limiting access. The research-preview framing buys time, not safety. If the recipe is as clean as described, trained on nothing but audio, built on hardware plenty of labs can rent, then others will follow, and not all of them will keep a waitlist. We have already written about how hard it is becoming to spot a convincing fake. Dessa moves that problem from recorded clips to the live call.

What I find genuinely interesting, set against the fraud worry, is the bet underneath it. Block is wagering that the way to make machines talk well is to stop teaching them to read first. If that holds up, it is a quiet rebuke to a decade of text-first design.