Block Dessa: what it is and how to get access
2026-10-24 · 4 min read
Dessa is a speech-to-speech voice model from Block, the company behind Square and Cash App. It listens and talks at the same time, it was trained only on recorded human conversation with no text at all, and Block says it runs locally on a MacBook Pro with an M4 Max. You can't use it yet: there is a public waitlist, and early access is going to a small group of AI safety researchers.
The launch posts
Jack Dorsey announced it on October 23. His post opens with "We believe voice is the most natural way to interact with intelligence" and links to Block's Dessa page.
Almost two hours earlier, the lab's lead, Ragavan Thurairatnam, posted a demo video. @RagavanThurair says Dessa is a full-duplex model "that you can run on consumer GPUs," and he calls it "a GPT3 moment for speech." That is a big claim from the person who built the thing, so treat it as a pitch for now, not a verdict.
The demo worth watching is the one from @blocks. An intern named Cara plays a guessing game with Dessa, and the model gets the first answer right away. The clip is short, and you learn more about the turn-taking from it than from Block's spec list.
What makes it different
Block says most voice AI today works as a pipeline: it turns your speech into text, send the text to a language model, and read the answer back with a synthetic voice. Block's argument is that the text step throws away how you said something. In the Dessa page's words: "The most powerful AIs in the world can't hear how you feel."
Dessa skips that step. According to Block's page:
- Audio goes in and audio comes out, with no text in between
- It was trained from scratch on human conversation only, with no pretrained models
- It learned language from "hundreds of times fewer words" than text models use, Block says, though it gives no dataset size
- It can interject and backchannel ("yeah!", "got it!") while you're still talking
- Training ran on one NVIDIA GB200 NVL72 system, and inference runs faster than real time on a MacBook Pro M4 Max or better
TechCrunch's hands-on write-up says the demos sound "strangely human," with laughs, pauses and even crying. Voice models like ElevenLabs v4 rely on tags to set emotion. Block says emotion is built into Dessa, but TechCrunch points out it isn't clear whether tags are steering it behind the scenes.
How to get access
Right now there are two routes, and neither gets you a working model today.
| Route | Who it's for | What Block asks for |
|---|---|---|
| Public waitlist | Anyone | Sign up on the Dessa page |
| Safety researcher early access | AI safety researchers | Institutional email, optional Google Scholar or blog link |
What you can do today is listen. The Dessa page hosts about two dozen recorded conversations, including one about remembering a movie, a bathroom disaster, a refund call, life advice, and a few clips of Dessa talking to OpenAI's GPT-Live-1. Block labels those last ones as involving a third-party model that hasn't endorsed Dessa. Start with the refund call and the expressive conversation clip if you want to know how this would sound on a business line.
Block hasn't published pricing, a model size, a license, or any word on open weights. Benchmarks are missing too, so every quality claim so far is Block's own.
Our read
Full-duplex voice is turning into a crowded field. OpenAI shipped it with GPT-Live, and ElevenLabs just launched Eleven Realtime for phone agents. Two things about Dessa are new. It was trained on speech alone, and it fits on a laptop. If the second one holds up, a voice agent could run on hardware a business already owns, with no per-minute cloud bill. TechCrunch notes Square already uses voice AI for ordering, so it's easy to guess where Block wants to take this.
Releasing it to safety researchers first is the right call. A voice that sounds tired when it says it's tired is also a voice that could fool someone on the other end of a call. Block calls today's emotion-deaf voice AI a safety problem, but a model that hears and performs emotion this well brings risks of its own.
What it means for a small business
Phone AI keeps getting cheaper and more human-sounding, and Dessa is one more sign of it. What hasn't changed is the work you do before turning anything on: which calls an agent can handle, which ones need a person, and how callers get told they're talking to a machine. Our comparison of answering services vs missed-call text-back is a good place to start that list. If you'd like help mapping it against how your calls actually come in, New Face Design's free process audit starts there.