01What the Dessa AI voice is and how it works
The Dessa AI voice is a speech model built by Jack Dorsey's company, Block. You talk, it talks back, and sound goes in and sound comes out.
The team calls it full duplex. That is a fancy word, but it just means it can listen and talk at the same time, like you do on a phone call with a friend. You do not wait for a beep.
Sometimes you talk over each other, and that is normal.
It laughs in the middle of a sentence, gasps and whispers under its breath, and it never read a single word. People are saying it sounds more human than anything they have heard.
The team called it a GPT-3 moment for speech. GPT-3 was the model that made many people say the text AI thing is real, and they think Dessa could be that moment for voice.
Jack Dorsey said Dessa is big, but also really small. It is small enough to run on a MacBook.
02The problem with how most voice AI works
Most voice AI today works in three steps. Julian Goldie lays them out.
- You talk, and it turns your voice into text.
- A text AI reads that text and writes an answer.
- It turns the answer back into a voice.
That sounds fine, but look at step one. When your voice becomes text, a lot gets lost. Your tone, your speed, a sigh, a laugh, the crack in your voice when you are upset.
The AI only gets the words.
Take the sentence, I didn't say she stole my code. Push harder on a different word each time and the meaning changes. The Dessa team uses that exact sentence to explain it.
Text cannot tell the versions apart, but your voice can.
So the most powerful AI can write you an essay in seconds, but it cannot hear how you feel. The team says that leads to a worse experience and can even be dangerous.
03Why timing and emotion matter for voice AI
The team points to a real danger. Picture someone in pain talking to a voice AI. Their voice might be shaking, but the AI only sees the words.
It keeps going cheerfully, like nothing is wrong. Dessa is a first step towards fixing that.
The second problem is timing. Old voice AI waits for you to stop talking, then it thinks, then it answers. That creates an awkward gap.
- You start talking again at the same time as it does.
- You both stop.
- It feels like a bad video call.
Dessa learned from real conversations. It learned when to jump in and when to just say got it and let you keep going. That little hmm you do while someone tells you a story is called back channeling, and Dessa does it.
Because it never turned your voice into text, it never threw anything away. It can pick up on emotion, sarcasm, laughing, a frustrated sigh, and a voice that is breaking because someone is sad.
04How the Dessa AI voice was trained without text
The first big difference is how it learned. Every big voice AI you have used was built on text somewhere along the way. Dessa was not.
As far as the team knows, it is the first AI you can have a realistic live conversation with that was trained only on human speech. Julian compares it to how we learn to talk when we are young. We do not read a book.
We learn by listening to the people around us, and Dessa learns the same way.
The training setup was small.
- The team trained the whole thing on one Nvidia GB200 NVL72, which is one big rack of computer chips.
- Big AI labs often use thousands of chips spread over huge data centers.
- Even though Dessa heard fewer words, every word was much heavier to learn from.
It is a tiny team and one computer rack, with no text at all, and the result is a voice AI people say sounds more human than anything they have heard.
05The Dessa numbers: speed, laptop use and team size
Julian runs through the numbers, and some of them are pretty cool.
- Faster than real time: It can think and talk faster than the speed of normal talking.
- Runs locally: It runs on a MacBook Pro with an M4 Max chip or better. That means on your own computer, not in a company's cloud.
- One rack: The whole model was trained on a single Nvidia GB200 NVL72.
- Small team: The research lists six people plus a few interns.
The team is a small lab inside Block. It was started by people from a Toronto deep learning company that was also called Dessa. They have been doing deep learning since 2012 and building AI voice models for over eight years.
They report straight to Jack.
A voice AI this good that runs on a laptop is a big deal, and the next section explains why.
06Why running the Dessa AI voice on a laptop matters
When a voice AI runs in the cloud, your voice gets sent to someone else's computer. Then you wait for the answer to come back.
When it runs on your own machine, your conversation can stay with you. There is also less delay, because the sound does not have to travel anywhere.
For a business, that is a privacy story as much as a speed story. If a client talks about their business on a call, keeping that on your own machine is a big plus.
Think about where voice already shows up in your day:
- Strategy calls.
- Client calls.
- Sales calls.
A voice AI that can hear emotion is a very different thing from one that only reads words. Voice is also how many customers still want to talk to you. They call, send voice notes on WhatsApp and jump on sales calls.
For a long time, putting an AI on the phone meant putting something robotic on the phone, and customers could tell straight away. That gap is closing fast.
07Dessa AI voice access, limits and safety
Voice AI is not new. You have had Siri and ChatGPT voice mode for years, and lots of companies sell AI phone agents. What is new is how Dessa is built and what it can hear.
You cannot plug Dessa into your business today. Block is giving early access to a small group of AI safety researchers first. They want outside people to find risks they have not spotted yet.
There is a public wait list as well.
Julian is open that it is not perfect and not fully out. It brings real questions, such as how to keep people safe when AI sounds as human as this. Block is being open about that, which is why safety researchers get it first.
The team gave three reasons for building it.
- Curiosity: could an AI learn to talk like a human with no written words at all? It can, but it is very hard.
- They believe voice is the most natural way to talk to any AI, and today's voice AI wastes it by squashing everything into text.
- Safety: AI is getting more powerful, and it still cannot hear the emotion in our voices.
08Who should care about the Dessa AI voice
If you run a business, you should care. The AI that talks to your customers is going to sound more like a person. Julian's view is that the businesses that understand this early will have a real edge.
As AI agents do more work for us, like answering calls and talking to customers, they need to know when someone is upset. That means hearing not just what was said, but how it was said.
Here are the kinds of jobs voice AI could help with once models like Dessa open up:
- Answering leads faster.
- Following up with people.
- Booking more calls without doing it all by hand.
For years, voice AI could talk but could not really listen. It heard your words and missed you. Dessa is the first big sign that this is changing: an AI that knows when to laugh, when to say yeah, and when to slow down because you sound sad.
