Block introduces Dessa, a human-like full-duplex voice AI model

Jack Dorsey’s company, Block, has unveiled Dessa, a new voice AI model designed to run locally on consumer hardware. Unlike many existing models that rely on text-to-speech generation, Dessa is a full-duplex speech-to-speech model trained entirely on human speech data.

The model is capable of simultaneous listening and speaking, allowing for natural, bidirectional conversations. Block researchers claim that Dessa incorporates human emotions, nuances, pauses, and cadences—such as laughter and sighs—directly into the model rather than using text-based emotional tags. The company stated that current AI interactions often feel unnatural because they convert voice to text, losing vital information, and they described the inability of powerful AI systems to perceive human emotion as ‘absurd’ and ‘potentially dangerous’ for user experience.