Block has released a voice model from its research lab that it says is the most authentic full duplex speech model yet built. The model, called Dessa, is trained fully on human speech, can talk and listen at the same time, and carries emotion, pauses, laughter and cadence in the generated voice itself rather than through text tags in a pipeline. TechCrunch reported the release on Saturday, following Block’s announcement on October 23.
Block’s argument is architectural: current voice AI converts speech to text and back, discarding most of what a voice carries, including how the speaker feels. Dessa’s knowledge comes from conversational data rather than a paired language model, and the model runs locally, on a consumer machine such as a MacBook M4 Max or better. There is no public access yet, and Block has put out a call for safety researchers to test the model first, an acknowledgement that a voice this convincing is also a fraud instrument.
The commercial route is visible without a launch. Block’s Square already uses voice AI to handle ordering for small businesses, and an on device model it owns would cut the cost and latency of those interactions while keeping customer audio off third party clouds.