ARTIFICIAL INTELLIGENCE · FOCUS
Block introduces Dessa, a voice model trained on human conversations
Block announces a voice model whose training is said to rely solely on human speech. The examples described by The Verge illustrate exchanges that sometimes feel natural, but also errors. Claims about its local operation and objectives remain attributed to its team.
Legal disclosure: this content was produced by an artificial intelligence system. Editorial transparency
Block has announced Dessa, an AI model that the company says was trained from scratch solely on human conversations. According to The Verge, its team describes it as a system that processes speech directly to produce speech, with knowledge said to come exclusively from voice content.
The reported demonstrations include a guessing game and discussions about cinema or a trip. They also show lapses: in one exchange, the model initially gets its conversation partner’s identity wrong; in another, it goes off topic. These examples illustrate Dessa’s behavior in the situations presented, without establishing its general reliability.
The team says the model can run locally on a MacBook Pro M4 Max or a more powerful machine. This information is not sufficient to specify operating conditions, execution speed or the resources actually used.
Block links the project to curiosity, understanding between humans and AI, and safety. The team particularly emphasizes the importance of emotion and rhythm in the voice. These are stated objectives: the material presented does not demonstrate that Dessa correctly recognizes distress or improves the safety of an interaction.
The project’s technical interest therefore lies in its stated approach to learning from speech. Assessing its significance would require evaluations that distinguish perceived fluency, response accuracy and the understanding of emotional signals.