Block's Dessa voice model runs locally, sounds human

Block, the financial services company led by Jack Dorsey, has released a voice model from its research lab called Dessa. Unlike many speech models that pair with a large language model for knowledge, Dessa is a full-duplex speech-to-speech model trained entirely on human speech, which Block says allows it to express emotion, pauses, and natural cadence without relying on text-based voice tags. The company argues that current voice AI loses information by converting speech to text first, leading to unnatural interactions. Block has not made Dessa publicly available; instead, it is inviting safety researchers to test it, likely to examine scenarios where a voice agent could deceive people. Given Block's business in finance and small-business solutions, a locally runnable voice model could eventually power more natural and lower-cost voice experiences across its products, including Square's ordering system.