Book a call
Block unveiled Dessa on 23 October, an AI voice trained only on human speech. It laughs, cuts in and hears tone. Here is what brands should do now.
On 23 October 2026, Block showed a new AI voice model called Dessa. Block is the company Jack Dorsey co-founded, and he shared it on X the same day (Source: Jack Dorsey on X, 2026 — Dessa post).
Dessa is a research preview, not a product. You cannot buy it or download it today.
But the clips Block posted are worth ten minutes of any brand team's time. In them, an AI laughs mid-sentence. It gasps. It says "yeah" while you are still talking. It finishes your sentence when you forget a name.
Block's point is that most voice bots do not sound like this. They wait for you to finish. Then they answer, often in a flat tone.
When this kind of voice reaches support lines, sales calls and voice ads, callers will hear an AI that sounds human. That is the part brands need to plan for.
One note before we start. Every claim and clip here comes from Block's own page. Block chose the recordings, and no outside lab has tested Dessa yet. Where something is Block's word, we say so.
Block calls Dessa "a true speech-to-speech model". In plain words, sound goes in and sound comes out (Source: Block, 2026 — Dessa).
Block says it trained Dessa "from scratch on nothing but human conversation". Its page adds, "No text, no pretrained models". The model learned to talk by listening, much like a baby does.
It is also "full-duplex". That means it can listen and talk at the same time, the way people do on a call.
Block's page has six main clips. A "More samples" list adds fifteen more (Source: Block, 2026 — Dessa). Block says each one is "a live recording of a person talking to Dessa". It also says, "Nobody steers it". The subtitles were added later.
Here is what the main clips show, in Block's own words.
That is a lot of human in one web page. But remember who picked the clips. Any company shows its best takes. Treat this as a demo reel, not a test.
Block shared a few facts about how Dessa was built. They are worth knowing, with one caveat. Nobody outside Block has checked them yet.
| What Block says | What it means for a brand |
|---|---|
| Trained on only one NVIDIA GB200 NVL72 | It did not take a giant cluster, by Block's account |
| Runs locally on a MacBook Pro M4 Max or better | A high-end laptop can run it, not only the cloud |
| Inference runs faster than real time | No awkward pause before it answers |
| Learned language from hundreds of times fewer words than text models | Speech alone was enough to learn to talk |
| Team has built speech models for over eight years | This is not a weekend project |
(Source: Block, 2026 — Dessa)
The laptop point is the one to watch. If a model like this runs on one machine, a brand may not need a big cloud contract to use it. That is still an if.
Block also explains why speech is hard to learn from. It says a single spoken word "takes thousands of times more bytes than the same word in text". And even one person never says a sentence exactly the same way twice.
To see why this matters, look at how a lot of voice bots work.
First, the bot turns your voice into text. Then a text model writes a reply. Then a voice engine reads that reply out loud.
Block's complaint is about the first step. It says current voice AI "turns your voice into text and throws away an incredible amount of information text alone can't carry" (Source: Block, 2026 — Dessa).
Think about what text drops. The sigh before a complaint. The joke told with a straight face. The shaky voice of someone who is upset.
Block gives a simple example. Say one sentence out loud, and stress a different word each time. The words stay the same. The meaning changes every time.
Block says Dessa "natively hears and speaks emotion, sarcasm, laughter, a frustrated sigh, a voice cracking with sadness". It says that if you run speech through text, "all of that is gone".
That is also why, in Block's view, most voice AI "feels stilted, slow and, worst of all, annoying". Those are strong words from a company with a model to show off. Still, anyone who has shouted "agent" at a phone menu will know the feeling.
Block's "More samples" list includes a clip called "Refund". It is the closest thing on the page to a real support call (Source: Block, 2026 — Dessa).
A person plays a customer. They say they ordered a bag, and it came with damage on the side. Nobody sets the scene for Dessa. It simply goes along and plays the shop.
Dessa sounds warm and calm the whole way through. It says "uh-huh" and "oh, yes" while the caller explains. That part is impressive.
Then the caller asks about a refund. Dessa says, "Yes, you can". A moment later it mentions an "option to refund partially". Then it says the caller "can get half of your money back" (Source: Block, 2026 — Dessa, Refund clip subtitles).
The caller pushes back, because the bag is brand new. Dessa replies, "I know the return policy makes no sense". Later it agrees to swap the bag with "no extra charges". Asked how long that takes, it says, "I can't give you an exact timeframe on that".
Be fair to Block here. Dessa was not connected to any real shop's rules. It is a research model, and Block says nobody steered it. It made the shop up as it went.
But that is the lesson. A voice that sounds this kind and this sure can still invent a policy. And callers tend to trust a warm voice more than a robotic one.
So the risk is not that AI voices sound fake. The risk is that they sound real while saying things your brand never agreed to.
Block's biggest reason for building Dessa is safety. Its page says, "Imagine being in serious pain while a voice AI has no idea and just keeps going" (Source: Block, 2026 — Dessa).
For a support line, that is a real upside. An AI that hears a frustrated sigh could slow down. It could say sorry sooner. It could pass the call to a person before the caller has to shout.
Block's life advice clip shows this soft side. Its labels on the clip read "Empathy", "Regret" and "Live your life". Block says the model "sounds genuinely empathetic" there.
Now flip it around. A voice that sounds caring is also a voice people trust. A sales bot that hears doubt in your voice could push at just the right moment. A bot that laughs at your joke starts to feel like a friend.
So the brand question is not only whether you can use this. It is also how your brand should sound when you do.
One clip on Block's page has no human voice in it at all. Dessa talks to GPT-Live-1, a voice model made by another company.
Block's labels on that clip read "Interruption", "Patience" and "Fast backchanneling" (Source: Block, 2026 — Dessa). Backchanneling means the small sounds people make while they listen. Think "yeah", "got it" and "ooh".
Block is careful here. It says both voices are AI. It also says the maker of GPT-Live-1 "is not affiliated with Block and has not reviewed or endorsed Dessa".
For brands, this clip hints at what may come. Your support bot may one day pick up a call from a customer's own AI assistant. Both voices would be AI, and both would sound human.
We looked at AI agents that handle support and sales in our piece on OpenAI Presence. The same questions apply here. Who is on the line, and who is in charge of what gets said?
The short answer is almost nobody.
Block has not published an app, an API or a price. Its page says its launch links, including a waitlist, are "coming soon" (Source: Block, 2026 — Dessa).
The one group getting in early is AI safety researchers. Block says it is "giving a small group of AI safety researchers early access to find the risks we haven't anticipated". There is an application form for them on the page.
The team behind Dessa is small. Block calls it "a (very) small AI research lab". It was founded by former members of Dessa, a deep learning startup from Toronto. The lab reports to Jack Dorsey, and the page is signed by Ragavan Thurairatnam, the Dessa lead.
Block says the team has built deep learning models since 2012. It has built generative speech models for over eight years (Source: Block, 2026 — Dessa).
So there is no tool to test this week, and no date to plan around. But there is time to get your own house in order.
You cannot buy Dessa. But you can get ready for voices like it, and that work is mostly writing.
Most brands have a tone guide for words on a page. Very few have one for a voice on a call.
A voice that can laugh, sigh and cut in needs rules. Here is what to put in yours.
The Refund clip is your test case. Play it to your head of support. Then ask what your own bot would have said.
When a vendor pitches you a human-sounding voice, run a few tests of your own. Interrupt it mid-sentence. Sigh, and see if it notices. Ask about a policy it should not know. Ask it straight out if it is a human.
The second document is shorter. It is one rule. Your AI voice always says it is an AI.
Write it down in plain words.
The law is moving this way too. The European Commission explains the rule in the EU AI Act. It says people using AI systems such as chatbots "should be made aware that they are interacting with a machine". It also says the Act's transparency rules come into effect in August 2026 (Source: European Commission, 2026 — AI Act).
If you sell into Europe, ask your lawyer how that applies to voice calls. If you sell elsewhere, the rule is still good manners. A caller who finds out late that the voice was an AI will feel tricked.
We made a similar point about marking AI content in our piece on Claude watermarks and C2PA. The tools change. The need to be honest about AI does not.
Three things are worth taking from this launch.
First, the gap between AI voices and human voices is closing, at least in demos. Block's clips show laughs, gasps and interruptions that most voice bots do not have.
Second, nothing here is ready to buy. Dessa is a research preview, and every claim is Block's own.
Third, the useful work is cheap and can start now. A voice guide and a disclosure rule take an afternoon to write. They will be ready when a vendor pitches you a voice that sounds human.
If you already run AI calling agents, like the ones in our hotel marketing guide, start there. Play the Refund clip to the team. Then check what your own bot says about refunds.
Dessa is a speech-to-speech AI model from a small research lab at Block. Block showed it on 23 October 2026 as a research preview. Block says it was trained only on human speech, with no text.
Not yet. Block has not published an app, an API or a price. Its page says launch links, including a waitlist, are coming soon. A small group of AI safety researchers gets early access.
Block says most voice AI turns speech into text first and loses tone. Dessa skips text. It listens and talks at the same time, says "yeah" while you talk, and laughs mid-sentence.
Block says it runs locally on a MacBook Pro M4 Max or better, faster than real time. No outside lab has tested that claim yet.
No. They are live recordings that Block chose and posted. Block says nobody steered the model. Treat them as a demo reel, not a benchmark.
Write a voice guide for AI calls and a rule that your AI always says it is an AI. Ban it from making up refunds, prices or delivery times.
It depends on where you sell. In the EU, the European Commission says people should know when they deal with AI systems like chatbots. Elsewhere, check your local rules. It is good practice either way.
