Block, the company founded by Jack Dorsey, has introduced Dessa, a new full-duplex voice model capable of local execution. The model is designed to simulate human emotional expressions, including laughter, pauses, and choking, by training exclusively on human speech data using a single rack of Nvidia GPUs.
Unlike many existing voice AI solutions that rely on converting speech to text, Dessa processes audio directly. This approach allows the model to capture nuances that are often lost in text-based pipelines. The model can adjust its tone, speed, and emotional state in real-time, as demonstrated in various interaction samples.
The model is optimized to run on consumer hardware, specifically requiring an Apple M4 Max chip or higher. While Block has not yet released the model for public use, it is currently inviting security researchers to test the system, focusing on potential risks related to AI-driven deception.
Block, which primarily provides financial services for small businesses, already utilizes voice AI for order processing through its Square platform. The development of this proprietary, edge-capable model aims to reduce operational costs and expand the company's voice-based product offerings for merchants and users.
