Block — the payments company behind Square and Cash App — unveiled Dessa on October 23, 2026, a full-duplex speech-to-speech model with an unusual training story: it was trained only on human conversation, with no text data and no pretrained models, according to Block. The result, in the demos TechCrunch highlighted, sounds eerily human — complete with the pauses, overlaps and interruptions of real talk.

Full-duplex matters here. Unlike turn-taking voice assistants that wait for you to finish, Dessa is designed to listen and speak at the same time, the way people do — talking over you, backchannelling, and recovering mid-sentence. That is the hard part of conversational voice AI, and it is why Block's approach is drawing attention beyond the usual voice-model launches.

There is a catch for anyone wanting to try it right away: Dessa is a research preview with a waitlist, and Block has published no benchmarks yet, so there is no independent score to compare against rival speech models. What Block does claim is practical: the model runs locally on a MacBook Pro with an M4 Max chip or better, putting a frontier-style voice model on a laptop rather than behind a cloud API.

For voice developers, Dessa is one to watch rather than one to build on today. If the local, conversation-only approach holds up outside demos, it points to a future where natural voice agents run on-device — private, low-latency, and free of cloud bills. The waitlist, not a benchmark table, is the next milestone.

Sources