Dessa takes a speech-native approach

Block has introduced Dessa, a conversation model it says was trained from scratch using only human speech, with no text or pretrained models. Unlike systems that turn speech into text before responding, Dessa is designed to take audio in and speak back, while listening and talking at the same time. Block says it can pick up cues such as emotion, sarcasm, laughter and pauses.

That matters for businesses considering voice AI. A spoken request can carry meaning in how it’s said, not just in the words themselves. Block argues that flattening speech into text can lose those cues, and says Dessa is intended to make conversation feel more natural. The company also says the model runs locally on a MacBook Pro M4 Max or better, and is inviting people to join a waitlist. These are claims from Block’s announcement, not evidence yet of how well the model performs in everyday business use. The Hacker News discussion drew 480 points and 180 comments, a sign that the speech-only approach has caught people’s attention.

For operators, the practical question is whether a model like this could handle real customer or staff conversations more usefully, especially when tone matters. That will depend on what becomes available and how it performs beyond the demos. This is the kind of development we build into an AI command centre, so teams can track what’s emerging and decide where it fits. Block’s announcement includes sample conversations and a call for safety researchers to examine potential risks.