Block launches Dessa voice model capable of running locally on Mac hardware

Block, the financial technology company led by Jack Dorsey, announced a new conversational model called Dessa that processes spoken sound directly. Most existing voice assistants first convert spoken words into text, stripping away vocal context. Block reported that Dessa was trained exclusively on raw audio recordings of human conversations rather than written datasets or existing text models.

Because the system relies on sound rather than text, it picks up on subtle cues such as laughter, emotional inflections, and pauses. The architecture supports bi-directional interaction, allowing the system to listen and respond simultaneously while interrupting or acknowledging the speaker naturally. Block demonstrated that the software can run locally without an internet connection on modern Apple laptop hardware.

Block stated that it is offering early access to safety researchers to identify risks specific to audio native artificial intelligence. What remains uncertain is how effectively a model trained on sound can retain complex factual information compared to massive text based models. Additionally, it is an open question whether Block will eventually integrate this lifelike voice system into its consumer financial apps or keep it as an independent research venture.