Block’s Dessa learned to talk by listening, and it sounds like it

Block has introduced Dessa, an AI model it says was trained from scratch on nothing but human conversation. It is a true speech-to-speech system, sound in and sound out with no detour through written text, and Block says it can run locally on a MacBook Pro M4 Max or better. Most voice assistants turn speech into text, reason in text and read an answer back, which loses timing, hesitation and emotion. Dessa bets that a model learning from speech alone can keep those signals. Block’s team lists three reasons: curiosity, better understanding between humans and AI, and safety. Its argument is that speech is the most natural interface to any intelligence and that converting it to text wastes it. Other labs are probing the same idea, such as Kyutai’s speech-native math model (Voice of Reason).

The method is more interesting than the demos. The team’s pitch is that babies learn to talk without seeing a written word, so why not a model; its answer is that it works but is very hard, and that candor is welcome. Many voice products sound polished because they are reading from a script. Dessa sounds like it is working things out. The Verge’s sample conversations show real back-and-forth with natural rhythm, and also strange moments: mistaking a speaker for someone named Ronaldo, wandering into Bryan Singer’s Superman films during a chat about Christopher Nolan, and answering news of a surprise Paris trip with what sounds like teary gratitude. Beside OpenAI’s GPT-Live-1, Dessa comes across as the more natural speaker. Voice executives have said the hard part now is understanding the listener and being honest that it is a machine (voice AI’s ChatGPT moment).

The safety argument deserves a hearing. Block’s team says AI systems keep gaining power while staying nearly deaf to emotion in our voices, and imagines a person in serious pain talking to a voice AI that just keeps going. A model that can hear distress could improve support lines, accessibility tools and any setting where people talk to machines when things go badly. Giving safety researchers early access instead of rushing a release is the right order.

The downside: a model built to sound emotionally present and read emotion back can also be very good at persuasion and attachment. Teary gratitude is touching in a demo; in products built to maximize engagement or sell, it is a lever. Regulators are already testing this line, as with Florida’s bid to stop ChatGPT acting like a person (Florida). Running locally helps privacy because conversations need not leave the device, but it does not settle what a friend-like voice should be allowed to do with that trust. What Block does commercially will decide which side of the line Dessa lands on. More in our AI coverage and generative AI coverage.