Jack Dorsey’s Block, the company behind Square and Cash App, has introduced Dessa, a new AI model built to hold conversations much like a person. Instead of relying on text processing like most voice assistants, Dessa has reportedly learned just by listening to human conversations.
Dessa is a full-duplex speech-to-speech model. This means it can listen and speak at the same time, so it can interrupt, acknowledge what someone says, and respond without waiting for the other person to finish.
Block says Dessa was trained from scratch using only recordings of human conversations. It reportedly does not use written text or pretrained language models, so its understanding of language, facts, and conversation comes only from audio.
The company says this method helps Dessa pick up on things traditional speech-to-text systems might miss, like sarcasm, hesitation, laughter, and changes in emotion. In demos on its website, the model laughs in the middle of a sentence, speaks louder when asked, and even finishes a sentence when someone forgets a movie title.
In a demo, Dessa talks with OpenAI’s GPT-Live-1 voice model. Block says these recordings show live interactions without anyone guiding the responses. Still, these are first party demos and should be viewed that way.
There’s also a hardware side to this. Block says Dessa was trained on a single Nvidia GB200 NVL72 system with 72 Blackwell GPUs. The finished model can reportedly run on a MacBook Pro with an Apple M4 Max chip or better. Block also says it runs faster than real time, so conversations don’t have noticeable pauses.
Dessa isn’t publicly available yet. Block is inviting a small group of AI safety researchers to test the model, and others are on a waitlist. The company hasn’t shared its parameter count, model weights, or standard benchmark results.