Block trained a voice model on nothing but recorded human conversation, with no text step in the middle. The demo audio is convincing, and that's about all anyone outside the company can check.
What Block Released With Dessa on October 23
Block published Dessa on October 23. The model comes from the company's own research team, and the announcement sits on a single page at block.xyz/dessa with demo clips and a waitlist.
There's no API, no weights and no price. A small group of AI safety researchers got early access. Everyone else fills in a form.
Jack Dorsey's company runs Square and Cash App, so this is a payments business releasing a speech model. That detail explains several design choices below.
The Speech-to-Speech Model That Never Reads Text
Most voice assistants run a three-step relay. Your audio becomes text, a language model writes a reply, and a text-to-speech engine performs it out loud. Every hop drops something, and the model in the middle never hears you.
Dessa skips the relay. Block trained it from scratch on human conversation, so speech goes in and speech comes out. The company's argument is blunt: today's voice AI compresses sound into words and loses what words can't hold.
That claim is easy to make and hard to check. No paper, no third-party evaluation, no benchmark table. Until someone outside Block runs this thing on their own hardware and publishes what they hear, "trained only on speech" describes a recipe rather than a result.
Full-Duplex Voice AI and the Trouble With Interrupting
Dessa listens while it talks. Block calls that full-duplex, meaning both directions stay open at once, so you can cut in mid-sentence and the model responds instead of finishing its paragraph.
That's a harder problem than taking turns. A turn-based assistant waits for silence, transcribes, then answers. A full-duplex voice model has to judge, in real time, whether the sound it just heard was a cough, a backchannel "mm-hm," or the first syllable of an interruption.
OpenAI shipped something similar this year with GPT-Live, and this is the part users notice first. Latency decides whether a voice agent feels alive or dead.
Block hasn't published interruption latency, so there's no number to compare.
Dessa on a MacBook: Block's Local AI Model Claim
Block says Dessa runs on a MacBook Pro with an M4 Max or better. That's odd for a model that sounds this human, since convincing speech systems usually live in a data center behind an API.
TechCrunch reported the training run used a single rack of NVIDIA GPUs. Treat that as media reporting rather than a formal company statement about the training setup. The serving target is still a laptop.
Why would a payments company care about local inference? Square already sells AI voice ordering to restaurants, and every call handled by a hosted model bills by the minute, which a Friday dinner rush multiplies across every table. A local AI model turns that bill into a hardware purchase, and for a shop taking a few hundred orders a day, the math flips.
It also means the model has to be small. Block hasn't said how small.
The Voice Cloning Risk Nobody Has Measured Yet
A voice this good is a fraud tool. Scam crews already clone family members from a few seconds of social audio, and a model that handles pauses, laughter and a cracking voice makes the fake much harder to catch on a phone call.
Block's answer is to give early access to AI safety researchers first. Read that as a test of whether a voice agent can fool the person on the other end of the line, before the public gets a copy.
Reasonable. Also unverifiable from outside. There's no published red-team report, no stated policy on impersonation, and no detail on how Block would mark or detect its own output.
That gap is the whole story right now. The company is asking for trust on the one axis where trust is hardest to earn.
What Dessa Sounds Like in Block's Demos
The clips on the page are short and deliberate. Someone says they're tired from work, and the voice does sound tired. Someone hears about a surprise trip and thanks the speaker with what sounds like a catch in the throat.
Those moments are where text-to-speech usually breaks. Written language has punctuation for exactly this reason. It can't encode a swallow.
One caveat before you get excited. Demo reels are the easiest thing in the world to put together. Nobody publishes the clip where the model laughs at a funeral.
Samples only.
Why Square's Voice Ordering Is the Real Audience
Follow the money and Dessa makes sense. Square's restaurant tools already take orders by voice, and those calls are short, repetitive and loud. Frying oil. A drive-thru speaker. A customer who changes their mind halfway through a sentence.
A phone agent that catches the hesitation in "uh, no, wait, the large one" is worth more than one that only reads a transcript. That's the commercial case for skipping text, and Block doesn't need to say it out loud.
The company's three stated reasons are curiosity, better human-machine understanding, and safety. Plausible ones. They're also the reasons any lab gives when it has nothing to sell yet.
What You Can Do With Dessa Right Now
Not much. You can listen to the samples and join the waitlist. If you're a voice-security researcher, you can apply for the early access program, which is the only path to touching the model.
That's it. No sandbox, no rate card, no published terms.
If you run a restaurant on Square, nothing changes this quarter. Watch for two things. First, whether Square starts routing real orders through Dessa, which would be the first public sign that local inference works at business volume. Second, whether any independent lab publishes a spoofing test.
Until one of those lands, Dessa is a strong demo from a company with a good reason to eventually ship it. Keep the skepticism, skip the hype.
