Block announces Dessa, a voice AI trained only on human speech

In short: Block, the company led by Jack Dorsey, announced a new voice AI model called Dessa and shared demo conversations to show how it talks.

What happened

Block announced an AI model named Dessa. Block says it trained Dessa “from scratch on nothing but human conversation,” meaning the system learned by listening to people talk, instead of learning from written text.

Block and Dessa’s lead, Ragavan Thurairatnam, posted example chats online. In one clip, the conversation sounds mostly natural, but the AI starts off by calling the speaker “Ronaldo,” even though that is not their name. The Verge described some of these moments as awkward and uncanny.

Block says Dessa is “speech-to-speech,” which means it listens to your voice and replies with a voice, without making you type. Block also says it is “full-duplex,” meaning it can handle back-and-forth talking more like a phone call, where interruptions and quick responses happen in real time (like two people talking, not taking turns with long pauses). Block claims Dessa can run on “consumer GPUs,” which are the powerful chips used in many gaming PCs.

The company also pointed to other examples on the Dessa website, including a voice guessing game, a discussion about filmmaker Christopher Nolan that goes on an unexpected tangent, and a conversation where Dessa responds with what the site describes as “teary gratitude.”

Why it matters

Voice AIs are getting closer to sounding like a real person in a casual conversation, which could make them easier to use for everyday tasks. At the same time, the demos show these systems can still misunderstand people or act strangely, which matters if you are trusting them in customer support, banking, or anything personal.

Source: The Verge AI