Block Unveils Dessa, An AI Speech Model Trained Solely on Human Speech

Block, the company behind Square and Cash App, introduced Dessa, an AI speech model trained entirely on human speech without any text data. Dessa processes audio directly, enabling real-time detection of emotions, sarcasm, laughter, and pauses, and supports full-duplex communication. It runs locally on a MacBook Pro M4 Max and was trained on a single NVIDIA GB200 NVL72. The model is currently available only for testing by AI security researchers.