Dessa AI is Block's full-duplex speech-to-speech model trained only on human conversation audio. No text. No pretrained models. It learned to talk the way you did, by listening.
📺 Watch: OpenAI's GPT-Live-1 Just Beat Grok Voice?!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
Block released it on 23 October 2026 on its page at block.xyz/dessa. Block is the company behind Square and Cash App, and it calls Dessa "the most authentic full-duplex speech model ever built". That's a big claim, so here's what it actually means, what the demos show, and what it means for you if you use AI agents and voice in your business.
Quick note before we start. I cover voice AI a lot on my channel, from GPT-Live-1 to Voxtral 3 to ElevenLabs. If you want to keep up with tools like this as they drop and actually put them to work, that's what the AI Profit Boardroom ($69/mo) is for.
What is Dessa AI?
Dessa is a conversation model. Audio goes in, audio comes out. There's no step in the middle where your voice gets turned into words on a screen.
Block trained it from scratch on nothing but human conversation audio. As far as Block knows, it's the world's first AI you can have a realistic live conversation with that was trained only from human speech.
Here's what Block says it can do:
- Listen and talk at the same time (that's the full-duplex part).
- Know when to cut in and when to backchannel with things like "yeah!", "got it!" or "ooh".
- Hear and speak emotion natively: sarcasm, laughter, a frustrated sigh, a voice cracking with sadness.
- Run inference faster than real time, so there's no awkward wait.
- Run locally on a MacBook Pro M4 Max or better.
How Dessa differs from text-based voice AI
Most voice AI works like this. You speak. It turns your speech into text. A text model reads that text and writes a reply. Then another system reads the reply out loud.
The problem is everything text can't carry. Tone. Timing. Emotion. The pause before you say something hard. Block's point is simple: run speech through text and all of that is gone. That's why so much voice AI feels slow and stilted.
Dessa skips the text step completely.
| What matters | Typical voice AI | Dessa |
|---|---|---|
| How it works | Speech to text, text model, text back to speech | Audio in, audio out |
| Training data | Built on text-trained models | Only human conversation audio, no text, no pretrained models |
| Tone and emotion | Lost when speech becomes text | Hears and speaks emotion, sarcasm, laughter, sighs |
| Turn-taking | Waits for you to finish | Full-duplex, listens and talks at the same time |
| Backchannelling | Rare | Says "yeah!", "got it!", "ooh" to show it's with you |
| Where it runs | Usually in the cloud | Locally on a MacBook Pro M4 Max or better |
One more detail I found wild. Block says Dessa learned language from hundreds of times fewer words than LLMs need. And it had to learn far more than language: accents, pauses, breathing, humour, inflection, when to cut in.
What full-duplex and backchannelling mean
Full-duplex means two-way at the same time. Think of a phone call versus a walkie-talkie. On a walkie-talkie, one person talks, says "over", and then the other person talks. On a phone call, you can both talk, interrupt, and react in the moment.
Most voice assistants are walkie-talkies. Dessa is built to be the phone call.
Backchannelling is the little stuff you say while someone else is talking. "Mm-hm." "Yeah." "Oh wow." You don't even notice you do it, but it tells the other person you're listening. Block says Dessa learned when to do this from real conversation, so it can show it's with you instead of sitting there silent until you stop.
Put those two together and you get something that sounds less like a machine taking turns and more like a person.
The Dessa demos
Block's page has a set of live recordings. Block says nobody steers it. Dessa picks what to say, when to say it, and how, down to its own voice and personality. Subtitles were added afterward.
Here's what the demos show:
- Talking about the release. It laughs the moment something is funny, mid-sentence. Ask it to speak up and it gets louder. Block says nobody taught it that.
- Remembering a movie. The person blanks on a movie title and Dessa finishes the sentence.
- An anniversary surprise. Dessa gasps "oh my god" under its breath, then answers with what sounds like teary gratitude.
- Expressive conversation. A big range in intonation and speed.
- Life advice. Here it sounds genuinely empathetic.
- Dessa and GPT-Live-1. Dessa talks with OpenAI's GPT-Live-1. Block notes the developer of GPT-Live-1 isn't affiliated with Block and hasn't reviewed or endorsed Dessa.
There are more samples on the page too, and Block says more recordings will be shared on X.
If you're wondering how this fits into your own setup, book a free AI strategy session and we'll map out where voice and agents make sense for you.
Running Dessa locally and how it was trained
Two hardware facts stand out.
First, running it. Block says Dessa runs locally on a MacBook Pro M4 Max or better. Jack Dorsey posted that it's "small enough to run on your macbook". Ragavan Thurairatnam, the Dessa lead, posted on X that you can run it on consumer GPUs.
Second, training. Block says it trained Dessa on only one NVIDIA GB200 NVL72. That's notable because speech is brutal data to learn from. A single spoken word takes thousands of times more bytes than the same word in text, and no two recordings of the same sentence are ever identical.
Block hasn't published a parameter count, a licence, or whether the weights will be open.
Who built Dessa
Dessa comes from a very small AI research lab at Block. The team was founded by former members of Dessa, the Toronto deep learning startup. They've built deep learning models since 2012 and generative speech models for over eight years.
The lab reports directly to Jack Dorsey. The lead is Ragavan Thurairatnam, who called it "a GPT3 moment for speech" in his X post.
Why Block built it, including safety
Block gives three reasons:
- Curiosity. Babies learn to talk without seeing a word written down. Block wanted to see if a model could too.
- Better understanding between humans and AI. Speech is the most natural way to talk to any intelligence, and today's voice AI flattens it into text.
- Safety. AI keeps getting more powerful, and most of it still can't hear the emotion in your voice. Block's example: imagine being in serious pain while a voice AI has no idea and just keeps going.
That third one matters. If AI is going to answer support calls, take bookings, or talk to people at their worst moments, it needs to hear how they feel, not just what they said.
How to get access to Dessa
Right now there's a waitlist. Go to block.xyz/dessa, click Join the waitlist, and follow the form. I've written up the steps in my Dessa waitlist guide.
Block is also giving a small group of AI safety researchers early access to find risks it hasn't anticipated, and asks researchers to get in touch. The page says "Launch links are coming soon." There's no public release date and no stated price yet.
What Dessa AI means for businesses using AI agents and voice
Voice is becoming one of the main ways people talk to AI agents. I run Hermes every day inside my Agent OS. For voice agents you can set up today, read my Hermes voice agent breakdown and the Hermes Apollo voice setup.
Here's where it stands. Dessa is a conversation model. Block's page doesn't say it does tool calls or connects to agent tools. So as far as the page shows, you can't plug it into your agent stack today.
But the direction is clear. Voice AI that hears tone, reacts in real time, and runs on a laptop changes what's possible for:
- Customer support that notices when someone is frustrated.
- Sales calls that don't feel like talking to a phone menu.
- Voice front-ends for agents, the kind of thing I explored in Hermes agent robot.
- Private, on-device voice for teams that don't want audio leaving the machine.
The smart move now is to get your agent setup solid so you're ready when models like this open up. My Agent OS guide walks through how I run agents day to day, and Goldie Bench shows how I test models before I trust them.
Inside the AI Profit Boardroom ($69/mo) you get the Agent OS, a 30-day AI agent roadmap, daily tutorials as new tools drop, and 4 coaching calls a week. It's 3,400+ members and we're rated 4.9 on Skool.
Want help working out where voice AI fits in your business? Grab a free AI strategy session.
FAQ
Is Dessa AI free?
Block hasn't stated a price. Right now access is through the waitlist at block.xyz/dessa, with early access going to a small group of AI safety researchers.
Can I run Dessa on my own computer?
Block says it runs locally on a MacBook Pro M4 Max or better. Ragavan Thurairatnam also posted that it runs on consumer GPUs. Launch links aren't live yet.
Does Dessa work with AI agents like Hermes?
Not as far as Block's page shows. Dessa is a conversation model and isn't stated to do tool calls or connect to agent tools.
What makes Dessa different from ChatGPT voice?
Dessa was trained only on human conversation audio, with no text and no pretrained models. It's full-duplex, backchannels, and hears emotion directly instead of converting your speech to text first.
Dessa AI is one to watch. If you want a plan for voice and agents in your business, book your free AI strategy session.
