Block's Dessa voice model laughs and sighs like a person, which is both its promise and its risk

Most voice assistants wait for you to finish, then answer in a flat voice. In demos this week, a new model from Block does something else. It laughs in the middle of a sentence, can say “yeah!” while you are still talking, and sounds tired when it says it is tired.

The Dessa voice model, revealed on 23 October by Block, the company behind Square and Cash App, learned to talk from recordings of human conversation alone. You cannot use it yet. But it shows where voice AI is heading, with real promise and a real worry about scams.

A Dessa voice model that learned to talk by listening

Most voice assistants turn your speech into text, let a text model write a reply, then read it aloud. Block says Dessa skips the text: it is a speech-to-speech AI, so sound goes in and sound comes out.

According to Block, the team trained Dessa from scratch on human conversation, with no text and no pretrained models, using a single Nvidia GB200 NVL72, one rack of AI chips. It is also “full duplex”: it listens and speaks at the same time, so it can cut in or say “got it!” the way people do.

Block says it runs on a MacBook Pro with an M4 Max chip or better, faster than real time, without the cloud. It also says Dessa hears and makes laughter, sighs and sarcasm on its own.

Some claims are Block’s alone for now. It says that, as far as it knows, Dessa is the first AI you can hold a realistic live conversation with that learned only from speech, and that it needed hundreds of times fewer words than text models. No paper, benchmark or model size has been published to check them.

The work comes from a very small lab inside Block, started by former members of Dessa, a Toronto deep learning startup that Square acquired in 2020. It reports to Jack Dorsey, Block’s co-founder.

Why Block is letting safety researchers in first

Block calls emotion a safety issue. “Imagine being in serious pain while a voice AI has no idea and just keeps going,” the company wrote on its Dessa page. A model that hears a shaking voice could, in principle, slow down or offer help.

The same skill cuts the other way. Some listeners were amazed; others, CoinDesk reported, warned that scam calls could become harder to spot. TechCrunch noted it is unclear whether hidden tags steer the emotion in the demos.

For now Dessa is a research preview: a small group of AI safety researchers get early access, and the public can join a waitlist. “A model that learned to talk from sound alone is new territory, and we want outside eyes on it,” said Ragavan Thurairatnam, who leads the Dessa lab. Block has not said how it will prevent impersonation, whether other languages will follow English, or what it will cost.

When you can try Dessa, and what to do before then

There is no Dessa voice model release date. Block has not said when the public gets access, or whether its weights will be open to download. Running AI on your own computer is moving fast, as Meta’s Muse Spark Open on two Mac desktops showed.

The more useful step today is about AI voice scams. Scammers will use human-sounding voices too. Agree a family “safe word” for emergency calls. If a caller asks for money or a code, hang up and call back on a number you already know. The same habit protects you from the scam emails that follow a data breach.

Dessa’s best idea needs its safety work to keep pace

On balance, Dessa is a meaningful step. Voice AI that cannot hear distress is a real gap, and Block is right to name it. A model that runs on a laptop could keep conversations private and cut costs; TechCrunch notes that Square already uses voice AI for ordering. And if one rack is enough to train it, smaller labs may follow.

The caution is just as real. Every clip so far was chosen by Block, and nobody outside has tested it. A voice that sounds human can comfort someone in distress, and it can just as easily disarm someone a scammer is calling.

Our reading: Block is starting in the right order, with safety researchers before a public app. The fair test is whether, before a wider release, it shares how Dessa was checked and how listeners will know it is an AI.

Watch for two things next: a paper or benchmark that lets others check Block’s claims, and a clear answer on disclosure. Until then, the demos are a reminder that a laugh and a sigh no longer prove there is a person on the line.