Jack Dorsey’s Block Introduces Dessa, an AI Voice Model That Learned by Listening, Not Reading
Block, the company founded by Jack Dorsey and best known for Square and Cash App, has introduced Dessa, a new artificial intelligence model built to make voice conversations feel more natural, immediate, and human-like.
What makes Dessa especially unusual is how it was trained. Unlike most AI voice assistants, which typically convert speech into text, process that text, and then generate a spoken reply, Dessa was reportedly trained entirely through audio. In other words, it learned language by listening to human conversations rather than reading written words or relying on a pretrained text-based language model.
Dessa is described as a full-duplex speech-to-speech AI model. That means it can listen and respond at the same time, much like people do in everyday conversations. Instead of waiting for a speaker to finish, it can react in real time, acknowledge what is being said, interrupt naturally, laugh, adjust its tone, and respond to emotional cues.
Block says this audio-first approach allows Dessa to pick up on subtle parts of communication that standard speech-to-text systems may miss. These include sarcasm, pauses, hesitation, laughter, shifts in emotion, and changes in voice intensity. In demonstrations shared by the company, Dessa can laugh during a sentence, raise its voice when asked to speak louder, and even help complete a thought when a person struggles to remember a movie title.
One demo reportedly shows Dessa speaking directly with another advanced AI voice system in a live conversation. Block says the interactions were not scripted and that the model’s responses were generated in real time. Still, since these demonstrations come directly from the company, they should be viewed as early showcases rather than independent proof of performance.
There is also a notable hardware detail behind the project. Block claims Dessa was trained on a single Nvidia GB200 NVL72 system, a powerful setup containing 72 Blackwell GPUs. Despite that high-end training hardware, the company says the resulting model can run locally on a MacBook Pro equipped with Apple’s M4 Max chip or better.
Block also claims Dessa’s responses can be generated faster than real time, which could help create smooth conversations without awkward delays. If accurate, that would make it a major step toward AI assistants that feel less like software tools and more like responsive conversational partners.
For now, Dessa is not available to the public. Block is opening access first to a limited group of AI safety researchers, while others can join a waitlist. The company has not yet revealed important technical details such as the model’s size, parameter count, training dataset scale, model weights, or standardized benchmark results.
Dessa arrives at a time when AI companies are racing to build more natural voice assistants. Its biggest distinction is its training method: learning directly from sound rather than text. If Block’s claims hold up under outside testing, Dessa could point toward a new generation of AI voice systems that understand not just what people say, but how they say it.
