I hit play on a thirty-second clip and paused, because my ears were lying to me. A voice that sounded weary, amused and impossibly alive was answering like a person—only it wasn’t. You will feel the same jolt when Dessa speaks for the first time.
I’ve been following Jack Dorsey’s post-Twitter moves for years: Bluesky gets the headlines, but Square—now Block, Inc.—is the quieter empire behind Cash App, Afterpay and the short-lived Tidal flirtation. You know the company: cards, sellers, payments. Now Block wants to be heard.
At a demo, an intern played a guessing game — and Dessa guessed the object on the first try
That clip, posted by Ragavan Thurairatnam, Dessa’s team lead, is useful because it reveals two things at once: the model’s uncanny fluency and its limits. The model doesn’t transcribe your words, then type an answer and read it back. Instead, Dessa uses a speech-to-speech architecture that processes audio and generates audio directly—so it can speak and listen simultaneously.
When Dessa guessed correctly on the first try twice in a row, it felt like watching someone intuit a thought. But it also kept guessing after being told it had won—an ordinary AI stumble that reminded me that human-sounding voice does not equal human understanding.
What is Dessa and how does it work?
Dessa is Block’s new voice model, trained “from scratch on nothing but human conversation,” according to Block’s announcement. That training choice is the hook: instead of teaching it via text-heavy corpora and separate text-to-speech layers, the team fed it spoken interaction so the model learned pauses, breaths, tone and the kinds of micro-timing that give speech its emotional weight.
On X, a clip sounded like an African American man — and replies flooded in
One upload shows Dessa with a cadence that many listeners read as an African American male voice. It says, “man, I’m doing good, breh. I ain’t gonna even lie. I’m tired,” and goes on about work, picking produce and an imagined newborn named Darius Johnson. The post exploded with reaction—because voice isn’t neutral.
That clip highlights the ethical crosshairs: convincing mimicry raises questions about consent, representation and stereotypes. Block gave a small group of AI safety researchers early access, and the team says they made Dessa for curiosity and to make models more sensitive to emotion. But curiosity and safety often tug in opposite directions.
How is Dessa different from ChatGPT’s voice?
Listen to the side-by-side demos: ChatGPT’s audio feels robotic by comparison. ChatGPT typically translates speech to text, generates a response in text, then synthesizes audio—that pipeline leaves pauses and a flattened cadence. Dessa’s single-stage speech loop produces natural hesitations, emphatic stress and the micro-pauses you expect in conversation. The result is a voice that lands with more human color.
I want to be blunt: hearing Dessa is like putting on a pair of well-worn headphones and finding the music is suddenly in the room. It can be comforting. It can also be dangerous, because tone can persuade when facts do not.
At demos and posts, empathy and tears were shown — yet the model still tripped on context
Other examples show empathy, the ability to sound tearful, even comic timing. Those moments drive the visceral reaction: this feels alive. But then you hear the model continue guessing, or invent personal details that don’t hold up, and you remember it’s pattern matching, not lived experience.
That oscillation between believability and glitch is where policy and product collide. If a voice can mimic cultural cadence convincingly, who sets the guardrails? Who gets to pick the model’s persona? You should be asking those questions now, before these voices saturate customer service lines, political ads or social feeds.
At the technical frontier, Block says Dessa runs locally on a powerful Mac — and they invited researchers in
Block claims Dessa can run locally on any MacBook Pro with an M4 Max or better. That matters because local execution limits server-side tracking and gives users more control—if Block keeps that promise. They’ve also shared the model with a small group of AI safety researchers, a nod to the reality that voice tech needs external scrutiny.
Between accessibility and risk, the company now holds a choice: publish a model that anyone can run, or keep it gated. Either path shifts the balance of power in voice tech.
Can I run Dessa on my MacBook Pro M4 Max?
According to Block, yes—you can run the model locally on a MacBook Pro M4 Max or better. That’s a practical advantage: local models reduce latency and keep audio off remote servers. But expectations should be tempered—running a model is one thing, governing how it’s used is another.
I’ve spent years watching technologies that sound human get woven into daily life—IVR lines, customer bots, political phone banks. Dessa is not the first to mimic us, but its fluency is a step-change. It hangs in a room like a live band on a small stage, and people listen differently to a voice that seems present.
If you care about how software will talk to you—and about how that speech will shape your choices—you should be paying attention to the demos, the safety notes, and who gets to release copies of the model. Will we update our ability to spot synthetic voices, or will these models rewrite what we accept as authentic?
Block says Dessa was made for curiosity and to improve how AI hears emotion. I’m curious too—but I’m also wary. When a synthetic speaker can mimic cadence and grief, who decides which voices get amplified, and who gets silenced?