Block's New AI Voice Laughs, Pauses and Even Cries, and It Runs on a Laptop

Jack Dorsey's company Block just showed off something that sounds a little unsettling at first. It's a voice model called Dessa that doesn't sound like a robot reading a script. It laughs, it pauses mid-thought, and it can even cry. People who've heard the demos keep using the same word to describe it: human.

What makes Dessa different is how it works. It's a "full-duplex" model, so it can listen and talk at the same time, the way a real conversation works. You don't have to wait for it to finish before you jump in. OpenAI launched something similar earlier this year with GPT-Live. Block is taking a different route, though.

Most voice assistants pair a speech system with a big language model that supplies the knowledge. Block says Dessa's knowledge comes from the conversation data it was trained on. That training was done entirely on human speech, using a single rack of NVIDIA GPUs.

The part that should get developers excited is that Dessa runs locally. Block says a consumer machine like a MacBook with an M4 Max chip or better can handle it, so you don't need the cloud or a data center.

You can't try it yet, though. Block hasn't made Dessa public. It's asking safety researchers to test it first, which makes sense. A voice this convincing could easily fool someone on a phone call.

For Block, there's a clear business reason. The company already sells tools to small businesses, and Square uses voice AI for ordering. A strong voice model that runs on the device could open up many more uses for merchants and everyday users.

So the obvious question for the rest of us is whether we're ready for AI voices we can't tell apart from real people.