Block introduces Dessa, a voice AI built from conversation data
Block announced Dessa, a speech-to-speech AI model trained solely on human conversation, and shared sample voice chats that sometimes sound awkward or uncanny. Dessa can run locally on a MacBook Pro M4 Max or better, according to the company. Block says it built the model for curiosity, improving human-AI understanding, and safety.
Block, led by Jack Dorsey, unveiled Dessa, a voice model developed using only recorded human dialogue rather than text. Demo clips show it handling a guessing game, discussing Christopher Nolan movies, and reacting to a surprise Paris trip; one exchange includes a mistaken name and an odd detour into Superman films.
Dessa is described as direct speech-to-speech and can operate locally on a MacBook Pro M4 Max or stronger hardware. Block cites curiosity, clearer human-AI understanding, and safety as motivations, and plans limited early access for AI safety researchers.
Dessa’s release could affect voice-assistant users, developers, and safety researchers by making local, conversation-trained voice systems more accessible and by drawing attention to emotional cues in speech. If such models become common, people may expect AI voices to sound more natural and to respond to distress more appropriately, while companies may face pressure to test for awkward or uncanny behavior. The limited researcher access may shape how safety concerns are studied before broader deployment.