Block has introduced Dessa, a new full-duplex voice model developed by the company's research lab. The model is trained entirely on human speech and is designed to run locally on consumer-grade hardware, such as a MacBook M4 Max or better.
Unlike many existing speech models that rely on text-to-speech pipelines or voice tags to simulate emotion, Dessa generates expressions like pauses, laughs, and cries directly from its speech-based training. The model is capable of simultaneous listening and speaking.
The company has not yet provided public access to the model. Block is currently inviting safety researchers to test Dessa, specifically to evaluate potential risks associated with voice agents that could deceive users.
Block intends to leverage this on-device technology to expand its existing voice AI applications, which are currently used by Square for order processing. The company aims to support a wider range of business and consumer experiences at a lower cost.
