Block's Dessa voice model is trained only on speech

Block published Dessa on Friday, 23 October 2026, a voice model the company says it trained from scratch on nothing but human conversation, with no text and no pretrained models.

Dessa lives at block.xyz/dessa, where Block calls it "the Conversation Model trained only on speech." The team describes Dessa as a true speech-to-speech system and full-duplex, so it listens and talks at the same time, closer to a phone call than a radio. Current voice AI, the team argues, flattens speech into text and throws away the emotion, the sarcasm and the pauses, which a performance needs.

For anyone cutting dialogue or lip-syncing a dub, that is the claim worth watching. Block says Dessa hears and speaks emotion natively and runs inference faster than real time, so replies arrive without the lag that makes a dubbed track feel stitched. Block also says Dessa runs locally on a MacBook Pro M4 Max or better.

All of that is Block's own description. The Verge senior reporter Jay Peters listened to the posted clips and wrote that the shared chats can be a little awkward and uncanny, including one where Dessa mistakes the speaker for someone named "Ronaldo." TheStreet writer Mehab Qureshi, who said she listened to every clip on the page, wrote that Dessa greets a man named Dheer as "Ronaldo" before being corrected.

Those recordings are what Block has to show. The page says each is a live conversation with a real person, that nobody steers the model, and that subtitles were added afterward. In one, Block says, Dessa laughs mid-sentence and asks a speaker to move closer to the microphone. Nobody outside Block has tested the model: the company says early access is going to a small group of AI safety researchers. The Indian Express notes the release carries no parameter count, no API, no price, no weights and no published evaluation.

That leaves Block's largest claim open. Block writes that "as far as we know, Dessa is the world's first AI you can have a realistic live conversation with that was trained only from human speech," and no independent party has confirmed it. Full-duplex voice is not new either: OpenAI's GPT-Live is built on a full-duplex architecture, and its GPT-Live-1 voice session reached the API at $0.05 a minute, according to the Coursiv blog. Whether a creator can use Dessa's output commercially is not established, and Block has published no license. The lab was founded by former members of Dessa, the Toronto deep-learning startup Block acquired in 2020, Cryptobriefing reports.

Dessa is not downloadable yet: the page takes an email for a waitlist, and the only route open right now is Block's form for AI safety researchers.