Block Releases Dessa: Highly Expressive Local Full-Duplex Voice AI
Jack Dorsey's Block has introduced Dessa, a full-duplex, highly expressive local voice AI model that natively captures human emotions like laughter, pauses, and fatigue. The model is capable of running locally on consumer hardware starting with Apple's M4 Max MacBook. Unlike traditional cascaded voice models that convert speech to text and back, Dessa handles audio natively with built-in emotional expression and full-duplex capabilities. This development could allow Block to deploy low-cost, natural conversational voice products across its merchant and consumer financial services. Dessa was trained exclusively on human voice data using a single Nvidia GPU rack, and its knowledge is derived directly from its conversation training data rather than an external large language model. It is currently not publicly available for testing, as Block is inviting security researchers to evaluate potential risks such as human deception.
## BACKGROUND
A full-duplex speech model is a real-time voice interaction system capable of both listening and speaking simultaneously, breaking away from traditional half-duplex turn-taking mechanisms. Block is a financial services company founded by Jack Dorsey that provides small business solutions and payment tools like Square.