Jack Dorsey’s Block has officially revealed an experimental AI model named Dessa, built from the ground up using nothing besides human speech data. Rather than relying on traditional text processing pipelines, the system functions as a true speech-to-speech model capable of executing local operations on high-end consumer hardware like an M4 Max MacBook Pro. To demonstrate the software's capabilities, the project lead Ragavan Thurairatnam and the engineering team published audio samples highlighting interactions that range from fluid dialogue to distinctly uncanny exchanges.
Unlike conventional conversational agents that translate voice queries into text before formulating a response, Dessa processes vocal inflection and pacing directly. Publicly shared transcripts and audio clips showcase the system engaging in unpredictable tangents—such as pivoting from a discussion about Christopher Nolan's filmography to unexpected remarks on Bryan Singer's superhero films—alongside erratic behaviors like misidentifying human participants. According to the development team, the architectural choice to train the model natively on vocal data stems from a desire to match how infants acquire language naturally without reading written text.
The Block team outlines three primary motivations behind the project: scientific curiosity, improving human-computer interaction, and addressing safety gaps inherent in modern voice interfaces. Standard voice AI systems often flatten human speech into text, discarding emotional context and failing to adapt to natural conversational timing. By prioritizing vocal emotion and local execution, the creators hope to build infrastructure capable of detecting nuanced human states, such as distress or pain. Access to the newly developed model remains heavily restricted, with the developers planning to distribute early builds exclusively to a select group of AI safety researchers.
Comments