Most voice assistants talk to you the way a call-centre IVR does. You speak, it waits, it thinks, it answers. Try cutting in mid-sentence and the whole thing stumbles. Jack Dorsey’s company Block thinks it has built something that behaves more like a real person on the other end of the line, awkward pauses, laughs and interruptions included.
On October 23, 2026, Block, the fintech firm behind Square and Cash App, unveiled Dessa, a voice AI model it says was trained from scratch on nothing but recordings of human conversation. No text, no borrowed chatbot underneath. It came out of a small research lab inside Block that reports directly to Dorsey, and for now only a small group of AI safety researchers will get early access.
Quick Facts
- Who: Block, the company led by Jack Dorsey (Square, Cash App)
- What: Dessa, a “full-duplex” speech-to-speech AI model that can listen and talk at the same time
- When: Announced October 23, 2026
- Where: Block’s official Dessa page and its X account; research lab inside Block
- Key claim: Learned language from “hundreds of times fewer words” than text-based AI models, and runs locally on a MacBook Pro with an M4 Max chip
What Is Block’s Dessa AI Voice Model?
Dessa is what engineers call a full-duplex speech model. In plain terms, it can hear you while it is speaking. So you can interrupt it, laugh, or throw in a quick “hmm” or “achha” the way you would with a friend, and it is supposed to react on the fly instead of waiting for you to finish.
Block says the model decides for itself what to say, when to jump in, and how to say it, including tone and volume. The company claims it can pick up on and express sarcasm, frustration, sadness and laughter, and that it can even laugh mid-sentence or speak louder when asked.
Dorsey himself called it the most authentic full-duplex speech model ever built. That is a big claim, and so far it rests mostly on Block’s own demo recordings.
How Dessa Is Different From ChatGPT-Style Voice Assistants
Most voice AI today works like a relay race. Your speech is turned into text, a text chatbot writes a reply, and another system reads that reply out loud. Every handover loses something. Your tone, your hesitation, the fact that you sounded upset: much of it disappears once everything becomes plain text.
Dessa skips the relay. It works directly on audio, so in theory the emotional cues stay intact. Block frames this partly as a safety issue, arguing that today’s AI often cannot hear distress in a person’s voice. Think of someone calling a helpline sounding shaky but saying “I’m fine”. A text pipeline only sees the words.
The training approach is also unusual. Block says Dessa never saw written text and was not built on top of an existing pretrained model. It was trained on a single Nvidia GB200 NVL72 system, which is serious hardware but modest next to the giant clusters used by the biggest AI labs.
Runs on a Laptop: Why That Matters
According to Block, Dessa runs locally on a MacBook Pro with Apple’s M4 Max chip or better, and faster than real time. That means no round trip to a cloud server for every sentence.
Local AI matters for two everyday reasons. First, privacy: your voice does not have to leave your device. Second, speed: on a patchy 4G connection, a cloud voice assistant can lag badly, while an on-device model does not care about your signal bars. For a country like India, where data quality swings wildly between a metro office and a moving train, that is not a small detail.
The catch: an M4 Max MacBook Pro is a premium machine, well beyond what most people carry. Phones are a different story entirely, and Block has said nothing about mobile.
The Team Behind Dessa and Block’s AI Ambitions
The model shares its name with Dessa, a Toronto deep learning startup. Former members of that company founded the small lab inside Block that built the model, and the lab answers directly to Dorsey.
Block is better known for payments than for AI research, so this is a notable side bet. A company that runs card readers for small shops and a popular money app now has its own homegrown voice model. Block has not said whether Dessa will ever show up inside Cash App or Square.
The Demos: Impressive, and a Little Weird
Early reactions to the recorded demos are mixed. Some moments sound strikingly natural, with laughter and quick back-and-forth. Others come across as stiff or odd. In one clip, the model confidently decided the human speaker was named Ronaldo, which he was not.
That is roughly what you would expect from a research preview. Human conversation is messy, and a model that learns only from that mess will pick up some of its strangeness too.
What We Don’t Know Yet
- How many hours of conversation audio Dessa was trained on, and where those recordings came from.
- How it scores on any benchmark. Block has published no paper and no third-party evaluation.
- Which languages it handles. There is no word yet on Hindi or other Indian languages.
- Whether there will be a public app, API or open weights, and when.
What This Means for You
Right now, nothing changes on your phone. You cannot download Dessa, and Block is only giving access to a small group of safety researchers. Applications go through the contact on the official Dessa page, and Block has said more recordings will be posted on its X account.
Still, it is a useful signal of where voice tech is heading. Expect future assistants, customer support bots and language tutors to sound less robotic and to handle interruptions better. A few practical points to keep in mind:
- As AI voices get more human, voice scam calls get harder to spot. If a “relative” or “bank official” calls asking for money or an OTP, hang up and call back on a number you already know.
- Ignore any app or website claiming to offer “Dessa AI” downloads. There is no public release.
- If you work in customer service or call-centre jobs, voice AI that handles real back-and-forth is worth watching closely.
Our Take
Training a voice model only on real conversation, then running it on a laptop, is a genuinely interesting idea, and Block deserves credit for trying something different from the “bolt a voice onto a chatbot” approach. But “most authentic ever” is marketing until outsiders can test it. No benchmarks, no paper and a Ronaldo mix-up mean we will believe the hype once researchers have had their turn.
FAQs
What is Dessa AI by Block?
Dessa is a speech-to-speech AI model from Jack Dorsey’s company Block. It can listen and speak at the same time and was trained only on recordings of human conversation.
Can I use Dessa right now?
No. It is a research preview, and early access is limited to a small group of AI safety researchers. There is no public app or API yet.
What does full-duplex speech model mean?
It means the AI can hear you while it is talking, so you can interrupt it or react mid-sentence, just like in a normal phone call.
Does Dessa need the internet to work?
Block says it can run locally on a MacBook Pro with an M4 Max chip or better, without cloud servers.
Does Dessa support Hindi?
Block has not said which languages Dessa supports, so there is no confirmation of Hindi or other Indian languages yet.
Based on reporting by The Verge, The Tribune (ANI) and AICoder, and Block’s official announcement.
Leave a Reply