Block's Dessa voice model laughs, pauses and cries - and runs on a laptop

Jack Dorsey's Block has unveiled a new voice model from its research lab, and it sounds unsettlingly human. The model, called Dessa, is trained entirely on human speech and a single NVIDIA GPU rack, and it runs locally.

Dessa is full-duplex, meaning it can talk and listen at the same time, much like OpenAI's GPT-Live model released earlier this year. Unlike typical speech systems paired with a large language model for knowledge, Block says Dessa draws its knowledge from the conversational data it was trained on.

Block argues that most voice models sound unnatural because text sits in the pipeline driving speech generation, and many rely on voice tags to inject emotion. The company claims emotion is built into Dessa, though it's unclear whether tags still steer it.

In a blog post, Block said interacting with current voice AI "can just plain suck," adding that it throws away information text can't carry. "The most powerful AIs in the world can't hear how you feel," the company said, calling that ridiculous, a worse experience and potentially dangerous.

Researchers posted sample conversations with different voices, cadences and emotions - in one, the bot said it was tired from work and sounded tired; in another, it sounded emotional, almost breaking down. Another clip showed a Block team member playing a game with the assistant.

Dessa can run on a consumer machine like a MacBook M4 Max or better. There's no public access yet; Block has called for safety researchers to test it, likely to probe cases where a voice agent could deceive people on the other end.

Block's core business is finance and selling solutions to small businesses, and Square already uses voice AI for ordering. An on-device model could power far more business and consumer experiences at lower cost.