Jack Dorsey’s Block made an AI model that talks to you like an awkward new friend

Technical Architecture & Context

Block has unveiled Dessa, a novel full-duplex speech-to-speech artificial intelligence architecture designed to emulate authentic human conversational patterns. Unlike traditional pipelines that chain automatic speech recognition, large language models, and text-to-speech synthesizers with inherent latency, Dessa is trained natively from scratch on raw conversational audio streams. This end-to-end approach allows the model to process acoustic nuances, pauses, and overlapping speech dynamically.

Systems Implications

Deploying a native full-duplex conversational model requires immense computational throughput at the inference layer. Managing bidirectional audio streams with minimal latency demands advanced edge-cloud coordination and optimized stream-processing frameworks. As conversational models become more human-like, the engineering focus shifts toward reducing inference overhead and eliminating unnatural conversational friction.

Primary Source & Reference

Original reporting and technical documentation: Jack Dorsey’s Block made an AI model that talks to you like an awkward new friend via The Verge (AI & Tech).