Block releases Dessa, a human-like, full-duplex voice model runnable on consumer GPUs

Block released Dessa, a full-duplex voice model trained entirely on human speech, capable of expressing emotions like laughs, pauses, and cries. Dessa can run locally on consumer GPUs such as a MacBook M4 Max or better, without needing a separate large language model for knowledge retrieval, deriving its knowledge from conversational data. While currently not publicly accessible, Block has invited safety researchers to test the model, noting that current voice AI often loses crucial information by converting voice to text. Square already utilizes voice AI for order handling, and this on-device model could lower costs for business and consumer experiences.