Block’s Human-Sounding AI Voice Raises a New Crypto Scam Question

Block has built a voice model that goes far beyond reading text out loud. It talks, listens, pauses, laughs and reacts with enough human cadence to make the old advice—“you can tell when it is a robot”—feel obsolete.

The model is called Dessa, and Block describes it as a full-duplex speech system built for the interruptions, pauses and overlapping turns that make real conversation difficult for conventional assistants. According to Block, it was trained directly on human speech and processes audio without first converting every exchange into text, allowing timing, tone and nonverbal cues to stay inside the model’s input.

That lets it listen while speaking, handle interruptions and respond to shifts in tone in real time, including laughter and surprise that would be flattened or delayed by a rigid transcription pipeline.

For a company built around payments, the breakthrough opens two very different doors: more natural tools for education, accessibility and customer support, and a much harder authentication problem for anyone moving money by phone or voice when cadence can no longer distinguish a person from software.

Block says it trained Dessa from scratch on a single Nvidia GB200 NVL72 computing system rather than adapting a text model and adding a synthetic voice layer. The finished model is compact enough to run locally on a MacBook Pro equipped with an M4 Max chip or better, a claim that points toward lower latency and deployment outside a large remote data center.

That local capability matters. A voice system that does not need to send every exchange to a remote server can respond faster and may offer more privacy.

It also lowers the infrastructure barrier for deploying sophisticated speech tools.

Block presented Dessa as a research preview rather than a finished consumer product. The company has not published a model size, benchmark table, pricing plan or public release date.

The early access group is deliberately narrow and includes AI safety researchers. That leaves the technical claims open to independent testing while limiting immediate exposure of a system designed to sound unusually human.

Block co-founder Jack Dorsey described the goal directly:

We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. https://t.co/r0qFLdvd6c

— jack (@jack) October 23, 2026

“Full duplex” is the key phrase. Most voice assistants take turns in a rigid pattern: listen, transcribe, generate, speak.

Dessa is designed to keep listening while it responds, closer to the overlapping rhythm of a real conversation.

A natural voice interface could make tutoring feel less mechanical, help people who struggle with screens and keyboards, and make customer-support systems less frustrating. Binance founder Changpeng Zhao quickly pointed to education as one possible use:

This might be very good for @GiggleAcademy. Teaching kids in current "AI voice" was a challenge. https://t.co/XrINEkqpos

— CZ 🔶 BNB (@cz_binance) October 24, 2026

The same qualities could make financial tools easier to use. A customer might talk naturally with a wallet or payment assistant instead of navigating menus, provided the system can prove what it heard, what it intends to do and who authorized the transaction.

That last condition is where the risk begins. Crypto users already face fake support agents, social-engineering calls and urgent requests to move funds.

A convincing voice can make those attacks feel personal even when the caller has never met the victim.

The answer is not to treat every advanced voice model as a fraud tool. It is to stop using voice familiarity as proof of identity.

Wallets, exchanges and payment services will need transaction-specific confirmation that is separate from the conversation itself: clear destination display, hardware-backed approval, withdrawal delays for changed addresses, and strong warnings when someone asks for a seed phrase or recovery code.

CoinDesk says Block has not disclosed Dessa’s model size, benchmark results or a public release date, so the demonstrations cannot yet be compared with competing systems on speed, accuracy or cost. The company is initially sharing the research preview with a small group of AI safety researchers rather than releasing it broadly.

The report also describes demonstrations in which Dessa laughed during a sentence and reacted to surprise with an emotional-sounding response. Block says those conversations were recorded live without a person directing the model’s answers, while the direct-audio architecture avoids the familiar speech-to-text-to-speech chain used by many assistants.

That cautious rollout is appropriate. The technology is impressive, but the financial world has to prepare for a simple new reality: hearing a trusted voice is no longer enough to trust the instruction.

Join the conversation!

We have no tolerance for comments containing violence, racism, profanity, vulgarity, doxing, or discourteous behavior. If a comment is spam, instead of replying to it please click the icon below and to the right of that comment. Thank you for partnering with us to maintain fruitful conversation.