Dessa is a speech model from a research lab at Block. It takes spoken audio as input and produces spoken replies, supporting real-time two-way conversation that keeps emotional cues because it was trained only on human voice recordings.