Block Releases Full-Duplex Voice Model Dessa, Trained Only on Voice Data

Techub News reports that Block's AI research lab has released the full-duplex voice model Dessa, which is trained solely on human conversational voice data without requiring text or pre-trained models. Dessa uses a voice-to-voice architecture, enabling simultaneous listening and speaking, supporting real-time interruption and mid-sentence laughter for natural interactions. The model was trained on a NVIDIA GB200 NVL72 rack containing 72 GPUs. During inference, it can run locally on a MacBook Pro M4 Max or higher-spec devices with speeds faster than real-time. Currently, Dessa is still in the research preview stage, and its weights, paper, API, or pricing have not been released.