Block Dessa Voice AI Preview: Access, Demos and Limits
Block Dessa is a new voice AI preview that can listen and speak at the same time. Block introduced it on October 23, 2026. The company has shared recorded conversations, but public access remains unavailable. For now, it is giving a small group of AI safety researchers early access.
The idea is easy to picture. In a normal conversation, you can interrupt, laugh or acknowledge a point before the other person finishes. A system that waits for a complete turn can miss that rhythm. Dessa aims to make those exchanges feel more natural.
How Block Dessa approaches conversation
On its official research page, Block says it trained Dessa from scratch using human conversation audio, without text or pretrained models. It describes a speech-to-speech system, rather than one that first turns your words into written text. Full duplex means it can keep listening while it talks.
Block also says the model runs locally on a MacBook Pro M4 Max or better. It says training used one NVIDIA GB200 NVL72 system. Those are company claims. They do not establish a memory requirement, a download you can install today or the hardware needed for a finished application.
What the preview does not prove
The recordings offer examples of the model’s timing and vocal range. However, Block selected those examples. A convincing clip cannot tell you how often a model gets facts right or handles an unexpected conversation well. That requires tests across many situations, including ordinary failures.
CoinDesk’s October 24 coverage confirms the preview and reports concerns about deception and scams. It also notes that Block has not disclosed model size, benchmark results or a release date. Therefore, a human-sounding voice should not be treated as evidence that the speaker is human.
For developers, the practical questions go beyond whether a demo sounds good. Can the system recover after an interruption? Does it make clear when it does not know an answer? What data would an application retain? Before building around a voice model, teams need clear answers and a way to test their own use cases.
For now, Dessa is research to watch, rather than a replacement you can download for an existing assistant. Local operation could be useful, but it does not by itself establish privacy, reliability or safety. Those depend on the eventual product and how people use it.
We have not tested Dessa ourselves. This report distinguishes Block’s claims from verified access. For more recent AI announcements and their practical limits, browse Dumb IT Dude’s Tech News archive.
Archive photograph of Jack Dorsey in May 2010. His company introduced the Block Dessa voice AI preview in October 2026. Photo: Joi / CC BY 2.0. Display crop; original photo license applies.
