Block Dessa: The AI Voice Model, and Who Can Use It Yet

Short answer: Dessa is a new AI voice model from Block, the company behind Square and Cash App, and almost nobody can use it yet. Block announced it on 23 October 2026. Early access has gone to a small group of AI safety researchers, and everyone else can join a waitlist. There is no app, no public demo you can talk to, no price and no release date.

This post is based on Block’s announcement page and the coverage that followed. We have not used the model, and neither has any independent tester we could find.

What Dessa is

Dessa is a speech-to-speech model. Block’s own summary is “audio in, audio out”. Most voice assistants today work in three steps: they turn your speech into text, send the text to a language model, and read the answer back with a synthetic voice. Dessa skips the text. It listens to sound and produces sound.

It is also full-duplex, which means it can listen and talk at the same time, the way people do on a phone call. Block says it learned when to interrupt and when to make the small “mm-hm” noises that tell a speaker you are following. Full-duplex models are not new as a research idea. PersonaPlex, published on arXiv in January 2026, is one recent example. What Block claims is different is how Dessa was trained.

What Block says is new about it

• —Training data: Block says Dessa was trained from scratch on human conversation only, with no text and no pretrained models underneath.
• —Efficiency: it says Dessa learned from “hundreds of times fewer words” than large language models, without giving a figure.
• —Hardware: training used a single NVIDIA GB200 NVL72 system, and Block says the model runs locally on a MacBook Pro with an M4 Max chip or better, faster than real time.
• —Expression: Block says it hears and produces emotion, sarcasm, laughter and sighs.

The work comes from a small research lab inside Block led by Ragavan Thurairatnam, who posted the launch and demo clips on X. Unite.AI reports that the lab was founded by former members of Dessa, the Toronto deep-learning start-up the model is named after, and that it reports directly to Jack Dorsey.

What is not proven

Every performance claim so far comes from Block. There is no technical paper, no published weights and no benchmark results. Block has not listed which languages Dessa speaks. The demo clips are described as live recordings, but nobody outside Block has said how they were chosen.

The clips themselves show both sides. Gizmodo’s write-up describes a guessing game in which Dessa got the answer first time, twice, and then kept guessing after it had already won. In another clip, a persona says it is tired from picking produce and then talks about a newborn son. Natural-sounding speech and consistent reasoning are different skills, and a voice that sounds human makes the second one easier to miss. That is the same habit worth keeping with any model: check the answer, not the confidence.

Why Block is starting with safety researchers

Block’s page argues that voice AI is “potentially dangerous” when it cannot hear the emotion in our voices, and calls a speech-only model “new territory”. It invites outside safety researchers to find the risks it has not anticipated. Researchers apply with a name and an institutional email from a university, company or lab.

The obvious risk is the one every realistic voice model raises: people may not be able to tell they are talking to a machine. We covered the same question for Gemini 3.8 TTS and its voice cloning controls, where Google added a watermark and a consent recording. Block has not described equivalent controls for Dessa yet. That is not a sign they will be missing, only that they are not documented today.

Can you use Dessa?

• —As a general user: no. You can join the waitlist on Block’s page. Block gives no date and no criteria for who comes off it.
• —As a researcher: you can apply for early access with an institutional email. Block calls the group small.
• —On your own computer: not yet. Running locally on a MacBook Pro is a claim about the model, not a download you can get. There are no weights to install.
• —In a country outside the US: Block names no country restrictions, but it names no general availability either, so the question does not arise yet.

One more thing to know before joining the list: the waitlist form collects your details under a privacy notice that Block dated 21 October 2026. If you have tried other speech tools, the questions are the same ones we asked of Falcon ASR and where its demo sends your audio: what is stored, where, and for how long.

What this means if you use voice AI at work

Nothing changes this week. The voice tools you can use today still work the three-step way, and the habits that make them useful still apply. The Verge’s report describes the same speech-to-speech design, and for now it is a research release, not a product. If it ships, the practical change is a voice assistant you can interrupt mid-sentence that answers in a tone, not just in words.

Two things are worth practising now regardless. First, treat a fluent voice as no more reliable than fluent text. Second, be clear about what you hand to an agent that can act for you, a question we worked through for Claude Duet and what it can reach. Realistic speech raises the stakes on both, as Gemini Live Avatar’s face and voice consent rules already show.

Coursium is a mobile app that teaches people to use AI at work. New models like Dessa will keep arriving, and most will be announced long before you can touch them. Knowing how to read a launch page, separate claims from evidence and check what a tool does with your data carries over to whichever one ships next. If that is what you want to practise, have a look at Coursium, or read why we built it.