Block introduces Dessa voice model with limited research access

Block has introduced Dessa, a speech-to-speech voice model that the company says can listen and speak simultaneously and was trained entirely on recorded human conversation. The model is available only to a limited group of AI safety researchers, with no public release date announced. Block announced Dessa on October 23. The company says it was trained from scratch without written text or a pretrained model, using an NVIDIA GB200 NVL72 system. Block describes the model as full-duplex: it can process incoming audio while producing spoken responses, rather than waiting for a user to finish before replying. The company says Dessa can run locally on a MacBook Pro M4 Max or better and perform faster than real time. Those performance claims come from Block; published reports provide no independent evaluation results. Block has not disclosed the model’s scale or the volume of its training corpus. Demonstrations released by Block show exchanges featuring pauses, laughter and changes in emotional tone, along with conversational interruptions. One demonstration included the model addressing a speaker as “Ronaldo,” an apparent mistake. The examples illustrate the behavior shown in the...