Interruptions, laughter and a change in speaking volume are part of a conversation, too. Block’s Dessa demo explores those details with a model the team says it trained from scratch on speech alone, without a pretrained text model.

It can listen and speak at the same time. The public recordings highlight timing, back-and-forth and vocal expression; they are selected demonstrations, rather than a published reliability benchmark.

This is a research preview, with limited early access for safety researchers. More conversation samples and the team’s account:
block.xyz/dessa