Block unveils AI model designed for conversational speech
Block has introduced Dessa, an AI model trained exclusively on human speech, aimed at enabling direct, full-duplex conversational interaction.
Financial technology firm Block has announced the development of Dessa, an AI model built to facilitate real-time, speech-to-speech communication. Unlike many generative AI systems that rely on text processing, the model was trained from scratch using a dataset comprised entirely of human conversation, intended to simulate the cadence and flow of natural dialogue.
Project lead Ragavan Thurairatnam has shared examples of the model in use, which demonstrate its ability to engage in back-and-forth communication. However, these demonstrations also highlight current limitations, as the system has occasionally produced incorrect contextual references or unnatural phrasing during testing.
This initiative marks Block’s entry into the competitive sector of speech-enabled AI interfaces. By prioritizing audio-based training, the company aims to move beyond standard text-based interactions, though the system remains in a development phase where accuracy and social nuance continue to be refined.