Financial technology conglomerate Block, Inc. unveiled Dessa on October 21, 2026, introducing what its research team describes as the world's first conversational intelligence trained exclusively on human sound. Developed by a dedicated laboratory comprised of veterans from Toronto deep learning startup Dessa—acquired by Block in 2020—the project operates under the direct supervision of Block Head Jack Dorsey. Unlike existing voice assistants that rely on chained pipelines converting audio to text, executing a text language model, and synthesizing synthetic speech, Dessa operates entirely in the continuous audio domain.
By discarding text intermediaries, the model preserves acoustic dimensions of human communication that traditional tokenizers erase. During live interactions, Dessa exhibits natural prosodic variation, adjusting pitch, cadence, and breath control in response to human vocal cues. It cuts in naturally during conversational pauses, laughs spontaneously mid-sentence when reacting to humor, and produces conversational backchannel acknowledgments such as 'yeah' and 'got it' while the counterparty continues speaking. The system also responds to vocal distress and emotional cracking with genuine acoustic empathy, addressing safety concerns where text-bound agents remain oblivious to physical pain or panic.
From an engineering perspective, the system represents an extraordinary achievement in sample and computational efficiency. While frontier transformer language models require tens of trillions of text tokens across multi-thousand-GPU clusters, Block trained Dessa from inception on a single NVIDIA GB200 NVL72 liquid-cooled rack. Despite the fact that raw uncompressed audio consumes thousands of times more storage bandwidth per word than ASCII text, the model assimilated conversational syntax, factual knowledge, and acoustic dynamics from hundreds of times fewer spoken words than typical web text corpora.
Inference performance has been optimized to execute locally without cloud connectivity. Block verified that Dessa runs faster than real time on high-end consumer hardware, including Apple laptops powered by M4 Max processors. To evaluate potential risks associated with an agent that acquires conversational agency purely through listening, Block lead Ragavan Thurairatnam announced that the laboratory is opening an early access sandbox to academic safety researchers, focusing on unintended cognitive biases and acoustic manipulation vectors before consumer deployment across Square and Cash App ecosystems.