Block research has announced Dessa, a full-duplex voice AI model trained directly on human speech rather than text pipelines.

Running locally on consumer hardware with an NVIDIA GPU, it supports simultaneous speaking and listening while incorporating natural vocal expressions.

How the new system works

The system's key difference is that it was trained exclusively on recordings of human speech, rather than text that is later converted to voice. According to the company, this allows it to sound unusually natural, including features such as pauses, laughter and crying.

The model has the ability to speak and listen simultaneously, meaning that the conversation can flow just like a normal conversation between two people.…