Block releases Dessa, an end-to-end full-duplex voice model

On October 24, Hacker News front page coverage said Block's dessa.com launched Dessa, which the company calls "the most realistic full-duplex voice model ever"; at the time the public page carried only a privacy notice, with no model parameters, benchmark scores or release method disclosed. Later the same day, The Verge reported the team says Dessa was trained from scratch on human conversation alone, handles speech input and output directly, draws all its knowledge from human speech, and runs locally on a MacBook Pro M4 Max or higher. Demo exchanges were mostly natural but included misidentifying a person and sudden topic drift. The team wants to improve human-AI understanding and address safety issues around recognizing emotion in speech, and plans early access for a small group of AI safety researchers.

Written by AI from the coverage · updated 1 hour ago

Not enough continuous observations to draw a trend yet.