Jack Dorsey’s Block introduces Dessa, a voice model that learns from speech

Block introduced Dessa on October 23. The company describes a model trained from scratch on human conversation, without text or pretrained models. Its full-duplex design lets it listen and speak at the same time, rather than waiting for a finished turn.

Block’s recordings demonstrate laughter, changing intonation, and responses to interruptions. The company says these are live conversations without someone directing the answers. These samples are demonstrations, not independent evidence that the system reliably understands emotions. The announcement presents Dessa as a research preview.

Roseram analysis: voice interfaces depend on timing as well as words. A useful evaluation should include overlapping speech, accents, background noise, and misunderstood instructions, alongside latency and naturalness.

A lifelike voice also makes disclosure and consent more consequential. Builders should make the artificial speaker identifiable and keep sensitive actions behind explicit confirmation. A compelling demo does not establish public availability or production reliability.