Block Launches Dessa: A Voice Model That Skips Text and Sounds Human Enough to Worry Fraud Watchers
An AI that speaks, and you can't tell it's an AI. In a demo that's called "impressive." On a phone call it's called "wait — I can't tell anymore."
Three key facts
Block unveiled Dessa, a voice model trained entirely on human speech rather than text. According to CoinDesk, unlike systems that convert speech to text and then generate a spoken reply, Dessa processes audio directly and can listen while speaking, so it handles interruptions, laughter, and shifts in tone. Jack Dorsey wrote on X: "Talking with it feels like talking with a person… full of emotion, nuance, pauses, and a natural cadence."
It's technically compact: trained from scratch on a single Nvidia GB200 NVL72 system, and small enough to run locally on a MacBook Pro with an M4 Max chip or better. Block says the demo conversations were recorded live without anyone directing the model's responses.
For now it's a research preview, not a public product. Block hasn't disclosed model size, benchmarks, or a release date; the first group to try it is a small set of AI safety researchers.
Wang Dou's Take
Let's separate "technically impressive" from "socially dangerous." Dessa's most praiseworthy quality is precisely its most dangerous one: it isn't "sounds synthesized-human," it genuinely processes human audio signals — interruptions, laughter, tone. Technically, that makes conversation natural. Socially, it makes "is the person on the phone real?" something you can no longer judge by ear.
Filmmaker Justine Bateman asked the sharpest question: "Why do AI voices need to sound human?" Her answer: "So you can trick people into lowering their guard." It's not a nice thing to hear, but it's the most honest footnote to this demo. A model's value isn't that it can deceive — it's that it makes being deceived cheap, requiring no technical skill at all.
But there's an angle where Block deserves credit: it gave the first version to AI safety researchers rather than opening it up immediately. That's uncommon in this industry — most companies ship first and think about safety later. The sequencing is worth remembering, even if it may also just be a presentable way of saying "productization isn't done yet."
The real question lies at the mechanism level. Voice was the last natural anti-counterfeiting layer for human identity. Once it falls, what does society use to verify "the other side is human"? The answer won't be "sound more human" — it'll be another mechanism entirely: voiceprints, digital signatures, trusted identity layers. None of which exist yet. Dessa isn't the start of the crisis. It just moved that day closer.
Source: CoinDesk · Block Dessa