Block unveils Dessa, a voice AI trained only on human speech, with no text at all

Block, the company behind Square and Cash App, on 23 October 2026 unveiled Dessa, a speech model that it says was trained from scratch only on recorded human conversation — no text and no pretrained language model.

What Block claims

• Dessa is speech-to-speech and full-duplex: it listens and talks at the same time, can be interrupted and makes small listening sounds such as “yeah” while you speak.
• Because it never converts speech to text, Block says it can hear and produce emotion, sarcasm, laughter and sighs that text-based voice assistants lose.
• Block says it trained the model on a single NVIDIA GB200 NVL72 rack and that it runs locally on a MacBook Pro M4 Max or better, faster than real time.

Caveats

All of this is Block’s own description, illustrated with recordings Block selected. According to TechCrunch and Block’s page, there is no public access, API or published benchmarks yet; Block is inviting AI safety researchers to test it first. TechCrunch notes that, unlike typical voice systems paired with a large language model, Dessa’s knowledge comes only from its conversational data — so how much it actually knows remains an open question. The Verge described some of the demo conversations as occasionally awkward and uncanny.