Block launches pure-speech-trained AI model Dessa for real-time emotion and sarcasm recognition

Block has launched the AI voice model Dessa, which it claims is the first to be trained entirely from human speech without any text data. It processes audio directly rather than converting speech to text first, enabling real-time detection of emotion, sarcasm, laughter, and pauses. Dessa supports full-duplex communication and can run locally on a MacBook Pro M4 Max; its training utilized a single NVIDIA GB200 NVL72 unit. The model remains unpublicized while the team is recruiting AI security researchers to test potential risks.