Block's Dessa trains a full-duplex voice model on speech alone

Block presented Dessa on 23 October 2026 as a research preview of a full-duplex speech model. Its official project page describes a system that listens and speaks at the same time, taking audio in and producing audio out. That is different from the familiar pipeline in which speech recognition first turns a person’s voice into text, a language model processes the text and a speech synthesiser produces a reply.

Block says Dessa was trained from human speech alone, without text or a pre-trained language model. The company argues that this lets the model learn parts of conversation that are hard to preserve in text, such as timing, intonation, laughter, hesitation and interruptions. It presents the model as able to react while someone is still talking, rather than waiting for a clean handover. Those are claims from the maker and have not yet been established by an independent benchmark.

The preview is unusually concrete about local hardware. Block says the model can run on a MacBook Pro with an M4 Max or better, which could make experimentation possible without sending every audio exchange to a cloud service. That does not mean the model is lightweight for every use: local speed, memory use and sustained performance still need independent testing.

Dessa is not a normal public release yet. Block has not published a general download, model card, benchmark suite or firm release date. The company says it plans to share an access link and give a small group of AI safety researchers early access to look for unexpected risks. Independent coverage likewise describes the demonstrations as company-selected examples rather than a complete evaluation.

For makers and businesses, the news points to a possible next step for voice agents: preserving tone and turn-taking instead of treating speech as a disposable input format. That could matter for accessibility, customer service, language practice and hands-free tools. It also raises practical questions about privacy, impersonation, consent and how an agent should respond when it detects emotion incorrectly. Dessa is therefore a notable research direction, not yet evidence that a production-ready voice platform has arrived.