Dessa: the AI voice that learnt to talk by listening

Every so often a demo lands that makes you stop what you are doing and play it again. Block's new voice model, Dessa, is one of those. Jack Dorsey shared it on X tonight, and having listened to the clips, it is the most convincingly human AI conversation I have heard. Not because the voice is polished, but because it behaves like a person: it laughs at the right moment, talks over you a little, and notices how you sound as well as what you say.

Here is what Dessa is, why it is genuinely different, and what it could mean for brands.

What is Dessa?

Dessa is a speech-to-speech AI model announced on 23 October 2026 by Block, the company behind Square and Cash App. It comes from a very small research lab inside Block, led by Ragavan Thurairatnam and founded by former members of Dessa, the Toronto deep learning startup. The team reports directly to Dorsey.

The headline features:

• Audio in, audio out. There is no text step anywhere in the loop.
• Full duplex. It listens and speaks at the same time, so it can interject, back-channel (the "mm-hm" and "right" we all use) and be interrupted naturally.
• Emotion aware. Block says it hears and produces laughter, sarcasm, sighs and sadness natively.
• Fast and local. It runs faster than real time and can run on a MacBook Pro M4 Max or better, so no awkward pause before it replies.

Why it is different from every voice assistant you have used

Most voice AI today is a relay race. Your speech is transcribed into text, a language model works out a reply in text, and a text-to-speech engine reads it back out. It works, but every handoff strips something out. Tone, timing, hesitation and emotion get flattened into words on a page, which is why most voice bots still feel like talking to a very polite form.

Dessa skips all of that. Block trained it from scratch on nothing but recordings of human conversation: no books, no web pages, no pretrained text model. It learnt to talk the way children do, by listening.

That is harder than it sounds. A spoken word carries thousands of times more data than its written version, even the same sentence said twice produces different data, and moving the stress in a sentence can change its meaning entirely. Despite that, Block says Dessa picked up language from hundreds of times fewer words than a typical large language model, and the whole thing was trained on a single NVIDIA GB200 NVL72 system. Block calls it "the most authentic full-duplex speech model ever built", and on the evidence of the demos that is not pure hype.

What the demos show

Block published a set of live recordings where nobody steers the model. Dessa chooses what to say, when to say it and how. Start with the launch demo Ragavan Thurairatnam, the Dessa lead, shared on X:

We made the first AI model you can have a realistic conversation with that was trained entirely on speech!

— Ragavan Thurairatnam (@RagavanThurair) 23 October 2026

Watch Ragavan Thurairatnam's Dessa launch demo on X

Block's Dessa page has 6 more recordings worth a listen:

• Talking about the release. It laughs mid-sentence the moment something is funny, and gets louder when asked to speak up.
• Remembering a movie. The speaker blanks on a film title and Dessa finishes the sentence for them.
• An anniversary surprise. It gasps under its breath, then replies with what Block describes as teary gratitude.
• Expressive conversation. A wide range of intonation and pace in a single exchange.
• Life advice. Block's example of how empathetic it can sound.
• Dessa in conversation with GPT-Live-1. 2 AIs talking to each other. Block notes that GPT-Live-1's developer is not affiliated with Block and has not reviewed or endorsed Dessa.

Listen to all 6 samples on Block's Dessa page

This is an early research preview, so expect rough edges. But the overall effect is a long way ahead of anything I have heard from a mainstream voice assistant.

Why marketers should pay attention

Dessa is not a product you can plug into your website yet. It is still worth taking seriously now, for 5 reasons.

1. Customer service gets emotional intelligence. Today's voice bots cannot tell a happy customer from a furious one. A model that hears frustration in someone's voice could change tone, slow down or hand over to a human at the right moment. Block cites exactly this as a safety reason for building it.
2. Brand voice becomes literal. Tone of voice guidelines have always been about words. When AI can control pace, warmth, humour and timing, brands will need to define how they sound, not just how they write.
3. Conversational commerce gets realistic. Voice-led shopping, booking and support have been held back by clunky, turn-by-turn interactions. Full duplex conversation removes much of that friction.
4. On-device changes the economics. A model that runs on a laptop rather than in a data centre points to cheaper, faster and more private voice experiences.
5. The bar for "good enough" just moved. Once customers hear what natural AI conversation sounds like, robotic IVR menus and stilted chatbots will feel even more dated.

The caveats

It is important to keep this in perspective.

• No public access yet. There is a waitlist, and a small group of AI safety researchers is getting early access first. There is no API or pricing.
• Knowledge is an open question. A model that learnt purely from speech may not know as many facts as a text-trained one. Block lists this as an open research question itself.
• Realism cuts both ways. A voice this convincing raises obvious questions about disclosure, consent and impersonation. Any brand using voice AI should be upfront with customers that they are talking to a machine.

What to do now

You do not need a voice AI strategy by Monday, but a few sensible steps now will put you ahead:

• Listen to the demo clips on Block's Dessa page and join the waitlist.
• Map where voice already touches your customer journey: phone lines, support, smart speakers, video and audio content.
• Start thinking about your brand's spoken personality alongside your written tone of voice.
• Agree a clear position on AI disclosure before you need one.

Voice has always been the most human way to communicate, and it has been the weakest link in AI. Dessa suggests that gap is closing faster than most of us expected.

If you would like to talk through what voice AI could mean for your brand, get in touch with the Clicky team.

Sources: Block, Dessa research preview · Ragavan Thurairatnam on X · Jack Dorsey on X