Block's Dessa Explained: The Voice AI Trained Only on Speech

Block just shipped a voice AI that was never shown a single word of text. Dessa, the research preview Jack Dorsey's company unveiled on October 23, 2026, learned to talk entirely by listening to recordings of people talking to each other. No transcripts, no pretrained language model underneath, no text data at any stage of training. And according to the people who've heard it, that's exactly why it sounds less like a chatbot reading subtitles out loud and more like an actual person with a bad day or a good joke.

If you build with voice AI, or you're trying to decide whether to bother, here's what Dessa actually is, what's verified versus hype, and why the "trained only on speech" part matters more than the demo clips suggest.

What is Block's Dessa?

Dessa is a full-duplex, speech-to-speech AI model built by a small research lab inside Block, Jack Dorsey's payments company (the one behind Square and Cash App). It takes audio in and puts audio out, with no transcription step anywhere in the pipeline.

The lab itself has a neat origin story: it's made up of former members of a Toronto deep learning startup that was also called Dessa, and the project reports directly to Dorsey through lead researcher Ragavan Thurairatnam. Thurairatnam called it "a GPT-3 moment for speech" in his own announcement on X, which is a big claim, but the training approach really is unusual.

Most voice assistants you've used, Siri, Alexa, ChatGPT's voice mode, work in three hops: speech gets converted to text, a language model reasons over that text, then a text-to-speech engine reads the answer back out loud. That pipeline throws away tone, timing, sarcasm, and the half-second pause before someone says something they don't want to say. Dessa skips all three hops and reasons directly over the audio waveform.

How was Dessa trained with no text data?

Block fed the model raw human conversation audio and nothing else. No transcripts, no books, no web text, no pretrained LLM backbone to fine-tune. The company says Dessa picked up language from "hundreds of times fewer words" than a typical large language model sees, because it's learning sound and cadence directly rather than learning from a mountain of text first.

The training run itself used a single NVIDIA GB200 NVL72 rack, a 72-GPU system that costs somewhere in the $4-8 million range depending on configuration. That's a real number worth sitting with: frontier LLMs train across thousands of GPUs and months of compute. Block says it did this on one rack. Whether that scales to something with broader knowledge is a separate question, but as a training-efficiency claim for a narrow domain, it's legitimate and worth taking seriously.

The payoff, per Block, is a model small enough to run locally on a MacBook Pro M4 Max, no cloud GPU required. That's the kind of distribution story that actually matters for developers, assuming the weights ever ship.

What does full-duplex actually mean here?

Full-duplex means Dessa listens and talks at the same time, the way two people on a phone call naturally overlap, interrupt, and backchannel with "yeah," "mm-hmm," or a laugh mid-sentence. Most voice AI today is half-duplex: it waits for you to stop talking, processes your turn, then responds. That turn-taking gap is why talking to Alexa still feels like talking to a machine even when the words are right.

Dorsey's own description, posted on X the day of launch: "Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence." Early listeners on X and Hacker News backed that up — multiple people said it's the most human-sounding AI voice they've heard, down to catching fatigue and sadness without anyone tagging the emotion for it.

That's the headline capability. It's also, so far, a demo capability. Nobody outside Block's own clips has used Dessa in an uncontrolled setting yet.

Is Dessa actually smart, or just good at sounding human?

This is the part the TechCrunch, the Decoder, and CoinDesk coverage all gloss over, and it's the one developers should care about most. Hacker News commenters who dug into the released samples flagged real limits:

• Dessa has no broad world knowledge and no tool-calling. It was never trained on text, so it has no native way to read a prompt, parse a JSON schema, or call an API.
• Demo transcripts include factual mistakes, like mixing up movie credits, and some exchanges read as filler, "speaking for speaking's sake," rather than substance.
• It only handles English right now.
• Because it skips text entirely, anyone who wants Dessa to actually do something, book a flight, answer a support ticket, pull up an order, will need to pair it with a separate LLM that handles reasoning and tool use while Dessa just handles the voice.

So the honest framing is: Dessa nails the acoustic and emotional layer of conversation and is years behind on the reasoning layer, because it was never built to have that layer at all. That's a deliberate trade, not an oversight, but it means Dessa today is a voice front-end, not a drop-in replacement for your existing voice stack's brain.

Is Dessa available to use or build with yet?

No. As of this writing there's no public API, no downloadable weights, no pricing, no benchmark numbers, and no published parameter count. Block opened a waitlist at block.xyz/dessa the same day it announced the model, and it's prioritizing AI safety researchers for the first round of access rather than developers or the general public.

If you want to build a voice agent today, Dessa isn't an option yet, full stop. It's worth watching, but don't block a roadmap on it.

Why is everyone worried about scams?

Because a voice model this convincing is also a voice model that's convincing when it's lying to your grandmother. CoinDesk's coverage captured the split reaction well: Binance's CZ Zhao floated it for children's education, while journalist Shiv Aroor warned "scam calls about to attain pandemic level." Filmmaker Justine Bateman pushed back harder, questioning why an AI voice needs human qualities like hesitation and emotion at all, given what that enables.

Block's answer, for now, is to keep the circle small: early access goes to safety researchers first, and the company hasn't disclosed model size or released open weights. That's a reasonable first move, but it's worth remembering that voice cloning and vishing scams were already a problem before Dessa existed. A model specifically tuned to nail human cadence and emotional nuance doesn't need to be malicious to make that problem worse. It just needs to exist and eventually leak, get reverse-engineered, or get a cheaper imitator.

How does Dessa compare to other voice AI models?

Kyutai's Moshi, released back in September 2024, already proved full-duplex, locally-runnable voice AI was possible. What Dessa adds is the "zero text" training claim and Block's size and backing, not the full-duplex idea itself. The Decoder's reporting makes the same point: sounding emotionally convincing is a different achievement than reasoning well, and nobody should confuse the two when reading the demo reactions.

What should developers actually take away from this?

Three things, honestly:

1. Don't build on Dessa yet. There's no API and no weights. Watchlist it, don't roadmap it.
2. The "skip the text bottleneck" idea is the real lesson, even if you can't use Dessa itself. If you're building voice agents today with something like OpenAI's Realtime API, Gemini Live, or a Kyutai Moshi-based stack, the emotional flatness users complain about traces back to that speech-to-text-to-speech pipeline. Pay attention to who closes that gap first with something you can actually ship.
3. Pair voice warmth with a real brain. Dessa's own limits, no tool calling, no world knowledge, are a preview of a pattern every speech-only model will hit. The practical architecture for the next year is probably a fast, emotionally expressive voice layer bolted onto a separate reasoning model, not one model doing both. If you're already composing agents this way, you're ahead of the curve; see our breakdown of Gemini 3.8 Live's extended thinking for how that split is starting to show up in production tools, and Mistral's Voxtral 3 for a text-capable voice model taking the opposite bet from Dessa.

If you're wiring voice into an agent pipeline and want the audio side to talk to your coding tools directly, it's also worth a look at Earshot's audio MCP server, which tackles a narrower but shippable version of the same "voice as an interface" problem Dessa is chasing at the research end.

Watch and read more