Block has introduced Dessa, an experimental speech-to-speech AI model trained from scratch on recorded human conversation rather than text. The system processes and produces audio directly, an approach meant to retain timing, emotion, pauses and other vocal signals that can disappear when speech is first transcribed.
Demonstrations show Dessa holding real-time conversations, playing a voice guessing game and responding with expressive delivery. The exchanges can sound natural, but they also expose rough edges: in one example the model incorrectly identifies a speaker as someone named Ronaldo, and in another it drifts into an unrelated film discussion.
The team says the model can run locally on a MacBook Pro with an M4 Max chip or better. Local operation could reduce latency and keep audio on the user’s machine, although Block has not released enough technical detail for independent evaluation.
Dessa is not currently available to the public, so its safety, reliability and performance claims remain difficult to test outside curated examples. The project nevertheless explores a different foundation for voice assistants: learning conversation from speech itself instead of building a text model with speech recognition and synthesis attached at either end.