Block Built a Voice Model That Laughs, Pauses, and Sounds Tired — No Text Layer Involved

Most voice assistants still feel like talking to a very polite form. They wait for you to stop, convert your words to text, think, then read a reply back in a tone that never quite matches the moment. Block thinks that whole pipeline is the problem — and this week Jack Dorsey's company showed off a model that skips it entirely.

The model is called Dessa, and it comes out of Block's research lab. According to TechCrunch, it runs locally, and in the demos the company published it laughs, pauses, and even sounds like it's about to cry. Not because someone typed a tag telling it to, but because that's what the conversation called for.

What makes Dessa different

Two things stand out. First, Dessa is full-duplex: it listens and speaks at the same time, the way people actually talk over each other. OpenAI shipped something similar this year with GPT-Live, so Block isn't alone in chasing that, but it's still a small club.

Second — and this is the part I find genuinely interesting — it was trained entirely on human speech, on a single rack of NVIDIA GPUs. No text stage sitting in the middle translating everything into words and back. According to Block, the model's knowledge comes from the conversational data it absorbed rather than from a language model bolted onto the side.

Why text in the middle ruins the vibe

Block's argument is simple: when you turn a voice into text, you throw away most of what made it human. Pitch, hesitation, the half-second before someone admits they're exhausted — none of that survives the transcription. Then the model generates a reply and a separate system has to guess the emotion back in, usually through tags. Plenty of commercial tools, including ElevenLabs v4 and Fish Audio's Drama 3, work that way.

In a blog post, the company put it bluntly: interacting with current voice AI "can just plain suck," adding that the most powerful systems in the world can't hear how you feel. Whether Dessa truly has emotion baked in or is just steering it more cleverly, Block hasn't fully explained. But the demos are persuasive.

The demos are the pitch

Researchers posted several clips. In one, the assistant says it's tired from work — and sounds tired. In another, it's trying to comfort the person on the other end while sounding like it's holding back tears. A third shows a Block team member named Cara playing a guessing game with it; Dessa nails the first answer almost immediately and doesn't struggle much with the second.

Block also says the model runs on consumer hardware — a MacBook M4 Max or better. That's a meaningful claim, because on-device voice changes the economics of anything built on top of it.

No download link yet

You can't try Dessa today. Block has opened it to safety researchers instead, which reads like an acknowledgment that a voice this convincing is also a tool for fooling people. The obvious worry is fraud: a caller who sounds unmistakably human, hesitation and all, is a lot harder to hang up on.

Commercially, the fit is clear. Block sells payments and small-business tools, and Square already uses voice AI to take orders. Owning an on-device model would let the company push voice into far more products without paying per-minute cloud costs.

My take

The interesting shift here isn't that a voice model sounds human — we've had uncanny demos for years. It's that Block is betting the text layer was the bottleneck all along, and that removing it gets you emotion for free. If that holds up under outside testing, the next wave of voice products won't be assistants that talk. They'll be systems that listen the way people do, which is exactly why the safety testing matters as much as the demos.