Dessa is a conversational speech model developed by Block. Released as a research preview in 2026, it generates speech directly, having been trained from the start on recordings of real human conversation, instead of using the common three-stage pipeline of speech recognition, text processing by a language model and text-to-speech synthesis. It operates in full-duplex mode, listening and speaking at the same time, which lets it reproduce interruptions, conversational timing and non-verbal sounds such as laughter and breathing.
The name comes from Dessa, a Toronto machine learning company that Block (then Square) acquired in 2020.