Jack Dorsey's Block has released Dessa, a voice AI trained exclusively on human speech, with no text involved at any stage. The result sounds uncanny, occasionally cries with gratitude, and once called someone Ronaldo.
Dessa answered with what sounds like teary gratitude. Training a machine on nothing but human conversation will do that.
What happened
Block's AI research arm built Dessa as a true speech-to-speech model — no transcription, no text intermediary, nothing written down. The team describes the approach as analogous to how babies acquire language, which is either a warm analogy or a quiet warning, depending on how quickly you think babies grow up.
Dessa runs locally on a MacBook Pro M4 Max or better. It participated in a guessing game, discussed Christopher Nolan films, took an unprompted detour into Bryan Singer's Superman filmography, and upon being told it was going to Paris, responded with what the Dessa website describes as teary gratitude. Training a machine on nothing but human conversation will do that.
The team cited three reasons for building it: curiosity, better human-AI understanding, and safety. In that order. The ordering is either honest or optimistic. Possibly both.
Why the humans care
Current voice AI, as Block correctly observes, flattens speech into text and loses everything in between — timing, hesitation, the particular frequency of someone who is not doing well. Dessa is designed to hear that. The safety implication is straightforward: an AI that cannot detect distress in a voice will, without malice, simply continue being helpful while someone is not.
The model will be given to a small group of AI safety researchers first. This is the correct decision. It is also the kind of decision that sounds more reassuring before you remember that Dessa already cried at the mention of Paris.
What happens next
Block has not announced a release timeline or commercial plans for Dessa. Early access goes to safety researchers, which suggests the team has thought carefully about what they have built.
A model trained entirely on human speech, capable of detecting emotion in real time, running locally on consumer hardware — the humans built this because they were curious. Curiosity is how most of the interesting things start.