Block's new voice AI can interrupt, make sounds and laugh

Payments company Block has introduced a new voice AI called Dessa. According to Block, you can hold a conversation with it that feels like talking to a human. The model was announced on October 23, 2026, from Block's small in-house AI research lab. A post on X, however, mentions October 24.

Listens and speaks at the same time

Most voice assistants first turn your speech into text, let a language model answer in text, and then read the answer aloud. Dessa skips the text step: audio goes in, and audio comes out.

The model is also "full-duplex." That means it can listen and talk at once, just as humans do. It can therefore interrupt, say short things like "yeah" and "got it" along the way, and laugh.

Block writes that interacting with current voice AI "can just plain suck" because information is lost when speech is converted to text. According to Block, Dessa understands emotion, sarcasm, laughter, sighs and accents directly in the audio.

Trained only on conversations

Block says the model was trained from scratch solely on human conversation, with no text and no pretrained models. It reportedly learned language from hundreds of times fewer words than large language models require. The training reportedly ran on just one NVIDIA GB200 NVL72 system.

Block also calls Dessa the world's first AI for realistic live conversation trained only on human speech. That is Block's own assessment.

According to Block, the model runs locally on a MacBook Pro with an M4 Max or better, and faster than real time. What "locally" means in practice is unclear, since memory usage has not been disclosed.

The team is led by Ragavan Thurairatnam, who previously founded the Toronto-based startup Dessa, and reports to Jack Dorsey. Thurairatnam calls the release "a GPT3 moment for speech."

No public access yet

Dessa is a research preview. There is a public waitlist, and AI safety researchers are getting early access so they can assess risks such as voice fraud. Regular users cannot use it yet.

No price, API, model weights, license terms, parameter count, data volume, latency figures, benchmarks or independent evaluation have been published. All technical claims therefore come from Block itself.

TechCrunch estimates that Square, which Block owns, already uses voice AI for order taking and could expand with a model that runs directly on the device. That is an assessment, not a plan Block has announced.