Mentioned in This Article
1. What is Block's Dessa voice AI?
Block, the payments company led by co-founder and chief executive Jack Dorsey (NYSE: XYZ), unveiled a voice artificial intelligence model called Dessa on October 24, 2026. In demonstrations released by the company the system sounds strikingly like a person, and the reaction online has divided between admiration for its conversational skill and warnings that it could make phone scams harder to spot. It is a research preview only, with no public release.
Dessa is described by Block as a model trained entirely on human speech rather than on written text. That distinction matters because most voice assistants in use today are built in stages: software first converts what a person says into text, a language model composes a written reply, and a separate system reads that reply aloud. Each hand-off strips out information that humans convey through sound, including hesitation, emphasis and mood.
Block says Dessa skips the text step and works on audio directly. It can listen while it is talking, which lets it cope with the messy features of real conversation: being interrupted mid-sentence, laughter, and shifts in tone. Those are exactly the moments at which older voice systems tend to stumble or fall silent.
2. How was Dessa built, and what hardware does it need?
According to Block, the model was trained from scratch on a single Nvidia GB200 NVL72 system, a rack-scale computing platform designed for large artificial intelligence workloads. Training from scratch means the company did not start from an existing third-party model and adapt it; it built the system on its own data and its own design.
The company also says Dessa is compact enough to run locally on a MacBook Pro fitted with an M4 Max chip or better. Local operation is notable for a speech model. Running on the user's own machine, rather than on a remote server, removes the network delay that can make spoken exchanges feel stilted, and it means audio does not have to be sent to a cloud service to be processed.
Block has not disclosed several details that technical observers would normally look for. The size of the model, measured in parameters, has not been given. No benchmark results have been published, and there is no release date. The audio samples that have circulated are demonstrations prepared by the company, not the product of independent testing, so claims about how human the system sounds rest for now on Block's own material.
That gap is common at this stage of a research preview. A company can show a model at its most polished long before it commits to numbers that others can check, and readers should keep the difference between a showcase and an evaluation in mind. Independent measurements of latency, accuracy across accents and behavior under poor audio conditions will be needed before the model's strengths and limits are clear.
3. What do the demonstrations show?
In one clip, Dessa laughs in the middle of a sentence. In another, it reacts to a surprise with what sounds like genuine emotion. Block says the conversations were recorded live and that nobody scripted or steered the model's replies as they happened.
Dorsey described the experience on X in personal terms. He wrote that talking with Dessa feels like talking with a person, and he pointed to its emotion, nuance, pauses and natural cadence as the qualities that set it apart. For a company known mostly for payments and financial software, the emphasis on emotional texture in a synthetic voice marks a notable departure from its usual public focus.
The clips spread quickly on social media. Because they are company-selected examples, they show what the system can do at its best rather than how it performs across a wide range of speakers, accents, noisy rooms and unpredictable requests. Until outside researchers can test the model under their own conditions, the demonstrations are best read as a statement of intent and capability, not a verified measure of performance.
The demonstrations also bring out why conversational timing is so hard to get right. People overlap, trail off, laugh at their own remarks and change direction without warning. A system that treats each turn as a clean block of text will miss these cues, while one that processes sound continuously has at least a chance of following them. Whether Dessa does so reliably outside curated clips is the central question for the research.
4. Why are users impressed?
Many of the early responses were admiring. An X user named Alexander Hoff said he had never heard anything artificial sound so human. Another, who posts as Hesamation, called the result straight-up science fiction. Their comments reflect a broader sense that synthetic speech has crossed from something clearly mechanical into something that can pass, at least briefly, as a person.
Changpeng Zhao, widely known as CZ and the founder of the cryptocurrency exchange Binance, also weighed in, and his interest was practical. He pointed to education, where natural voice interaction could help children learn through conversation. Zhao wrote that the technology might be very good for GiggleAcademy, his education project, and said that teaching children with the current generation of AI voices had been a challenge.
The comment hints at why lifelike speech is commercially attractive. Young learners respond to tone and warmth, and a voice that sounds flat or robotic can lose their attention quickly. A system that can adjust its delivery, react to interruptions and sound engaged is the kind of tool that education developers, customer-service operators and companion-app makers have long wanted.
5. Could a human-sounding AI voice fuel scams?
The same realism that impressed some listeners alarmed others, who see a tool for deception. Filmmaker Justine Bateman questioned why AI voices need to sound human at all. She argued that the realism can be used to mislead people, writing that it allows them to be tricked into lowering their guard.
Journalist Shiv Aroor raised a related concern about fraud, writing that scam calls were about to reach pandemic-level volume. His worry points to a familiar pattern: voice fraud depends on persuading a victim that the caller is who they claim to be, and a convincing synthetic voice can lower the effort and cost of doing so across many calls at once.
These are opinions from commentators, not findings from a study, and Block has not claimed that Dessa is suited to impersonation or that it is being released in a way that would enable it. Still, the debate echoes an argument that has followed each advance in synthetic media. Better realism widens the legitimate uses, such as accessibility, tutoring and customer support, while also lowering the barrier for abuse. For readers in finance and cryptocurrency, where impersonation scams already cause losses, the concern is especially close to home.
Defenses against voice fraud generally rely on habits rather than technology: calling a person back on a number already known to be theirs, agreeing a private code word with relatives, and treating any urgent request for money or credentials with suspicion however familiar the voice sounds. As synthetic speech improves, such precautions matter more, because the sound of a voice becomes a weaker proof of identity than it once was.
6. Who is behind the project at Block?
Block runs several businesses: Square, the Cash App payments service, the Bitkey bitcoin wallet, and Proto, its unit that makes bitcoin mining hardware. Its artificial intelligence research lab reports directly to Dorsey, which signals how seriously the founder treats the field.
The lab includes former staff of Dessa, a Toronto-based artificial intelligence startup that Square acquired in 2020. The model's name appears to carry that lineage forward, linking the new research to a team that joined the company several years ago.
The link between a payments firm and voice research may look unusual, but speech is a natural interface for financial services. Customers already call support lines, confirm transactions and ask questions about their accounts, and a voice system that understands emotional cues could in principle handle those interactions with more sensitivity. Block has not described any specific product plans for Dessa, however, and the research preview is separate from its commercial services.
The origin of the name also gives some context. The earlier Toronto startup was acquired by Square in 2020, and its alumni now sit inside the research group that answers to the chief executive. Continuity of that kind suggests the voice work grows out of years of accumulated expertise rather than a sudden pivot.
7. What happens next for Dessa?
For now, Dessa is not available to the public. Block is offering early access to a small group of AI safety researchers, who are assessing risks the company may have overlooked. That staged approach lets outside experts probe the model for misuse before any wider rollout is considered.
Block also argues that speech-based AI could eventually get better at recognizing emotional cues such as distress and pain, which text transcripts can miss. In that view, a model that listens to the sound of a voice, not only the words, could notice when a caller is upset or in trouble and respond accordingly.
The open questions are the ones the company has not yet answered: how large the model is, how it performs on independent benchmarks, when or whether it will be released, and what safeguards would accompany a release. Those answers will largely determine whether the enthusiasm or the concern proves more justified. Block's shares were trading at $76.45 when the market closed on the day of the announcement.