01What the Dessa AI voice model is

The Dessa AI voice model is a new voice AI from Block that aims to sound very human. It laughs, sighs, says yeah and got it while you talk, and you can cut it off mid-sentence like a real person.

It comes from a small AI research lab inside Block, Jack Dorsey's company. Dessa itself is a deep learning startup from Toronto. They have been building deep learning models since 2012 and speech models for over eight years.

Block calls it a full duplex speech model and says it is the most authentic one ever built. The wild part, as Julian Goldie points out, is that it was trained only on speech, with no text at all.

02What full duplex means for a voice model

Full duplex is the big idea here. Most voice AI works like a walkie talkie. You talk, you stop, it thinks, then it talks back, one person at a time.

Full duplex is more like a phone call. The model listens and talks at the same time. While you explain something, it can say hmm or yeah to show it is following.

If you interrupt, it stops and listens, and it knows when it is its turn to jump in.

  • Walkie talkie style: one side at a time
  • Full duplex: both sides at once

That is how real conversations work, and it is why most voice assistants still feel a bit like talking to a machine.

03Trained on speech only, with no text

The second big idea is how it was trained. Most voice AI starts with text. A model sits underneath, and the voice gets wrapped around it.

The Dessa AI voice model does not do that.

Block says it was trained from scratch on human conversation, with no text and no pre-trained models. When you turn speech into text, you lose a lot, such as tone, pauses, a little laugh or sarcasm. Dessa only learned from sound, so it picks up all of that.

Block also says it learned language from hundreds of times fewer words than a normal language model.

Older voice assistants turn your voice into text, think in text and read the answer back out loud. Dessa skips the text completely and goes from speech to speech.

04What the Dessa AI voice model can do

Julian describes what Block shows on its page.

  • Emotion: it hears emotion and shows emotion back, including sarcasm, laughter, sighs and even a voice cracking
  • Back channeling: the little yeah and got it sounds we all make when someone is talking to us
  • Laughing mid-sentence: it can laugh in the middle of what it says
  • Volume: if you ask it to speak up, it actually gets louder
  • Interruptions: you can cut it off and it stops to listen

Together, full duplex and speech-only training are why the demos sound so natural. Julian invites listeners to hear the demos and tell him in the comments whether they could tell it was AI.

Put simply, there are two changes at once. The model is full duplex, so it can listen and talk together instead of taking turns. And it goes from speech to speech without text in the middle, so the feeling is not lost.

Julian says it is the combination that makes the demos sound so natural.

05Dessa runs on your own computer

One part Julian likes most is that Block says Dessa can run locally. It runs on a MacBook Pro with an M4 chip or better, and it runs faster than real time. That means no awkward wait while it thinks.

This matters because most realistic voice AI runs in the cloud. You send your voice to a server and wait for it to come back. If a model like this runs on your own machine, your conversations can stay on your computer and there is no lag from a round trip to the cloud.

Block also trained it on surprisingly little hardware. It used one NVIDIA GB200 system for a model at this level, which Julian calls a very small setup.

Put together, the local running and the small training setup show how little hardware this kind of model may need, though Block has not said when you will be able to run it yourself.

06How to get access to the Dessa AI voice model

Right now access is by wait list only. Here is what to do.

  1. Go to block.xyz/desa.
  2. Join the wait list.
  3. If you are an AI safety researcher, apply for early access with your institution's email.
  4. Watch for launch links, which the page says are coming soon.

Early access is going to a small group of AI safety researchers first. Block says it wants outside researchers to find risks it has not thought of yet.

There is no download yet, and Block has not said whether it will be open source. Julian says he will keep an eye on it and test it properly when it opens up.

07Why voice AI matters for AI agents

Julian thinks voice is about to become the main way we talk to AI agents. Most of us type now, but typing is slow. Talking is how we already work with people.

The problem has been that talking to AI feels stiff. You wait, it talks over you, or it does not stop when you cut in, and it does not pick up on how you feel. Block also says voice AI that cannot pick up on emotion is a safety problem, and Dessa is a step toward fixing it.

Imagine calling your agent while you walk around. It hears that you are stressed or rushing and adjusts. That is where this is heading.

If you want help putting AI agents to work in your business, voice included, Julian points to the AI Profit Boardroom.