Jack Dorsey’s Fintech and Bitcoin focused company Block (NYSE: XYZ) has released Dessa, a research preview of a speech model that produces conversation so lifelike it has prompted both admiration and warnings about potential misuse.
The work comes from a compact internal lab that answers directly to Dorsey and builds on years of generative audio research.
Unlike conventional voice systems, Dessa never sees written text.
It was trained entirely from recordings of people talking, with no pretrained language models underneath.
Audio enters and audio leaves; the system learns language, timing, and expression solely by listening.
The architecture is full-duplex, meaning the model continues to hear while it is speaking.
This lets it insert short acknowledgments, finish a half-remembered phrase, or adjust its delivery without waiting for a clean turn.
Company recordings show it laughing the instant something becomes funny rather than after a prompt ends, raising its volume when asked to speak louder, and producing what sounds like a quiet gasp followed by near-tears when told of an unexpected anniversary trip.
Inference runs faster than real time, so there is no noticeable lag.
The finished model is small enough to run on a MacBook Pro equipped with an M4 Max chip or higher.
Training itself required only a single NVIDIA GB200 NVL72 rack.
Block’s researchers argue that most existing voice AI discards the very signals that make human talk meaningful—hesitation, sarcasm, a sigh of frustration, a voice that cracks.
By staying in the audio domain, Dessa retains those cues.
The lab frames the project as both an experiment in whether a machine can acquire spoken language the way an infant does and a practical step toward interfaces that notice emotional state.
They note that an AI that cannot hear distress while a user is in pain remains a safety risk.
Early access is therefore being offered first to a limited set of external safety researchers who will look for failure modes the internal team has not yet considered.
A public waitlist is open, but no weights, API, or commercial timeline has been published.
Listeners who heard the demos reacted in sharply different ways.
Some described the voice as the most convincingly human synthetic speech they had encountered, calling the exchanges science-fiction level.
Others focused on the downside.
Once a synthetic voice can interrupt naturally, match emotional tone, and sound spontaneous, the usual auditory tells that once flagged automated calls become unreliable.
Critics pointed out that the same realism that creates warmth can also lower a listener’s guard during a fraudulent call.
Changpeng Zhao, founder of Binance, took a more constructive view.
He observed that the current generation of AI voices has made teaching children through conversation difficult and suggested that a more natural system could be useful for educational projects such as his own GiggleAcademy, where spoken interaction is central to learning.
Block has released no independent benchmarks, parameter counts, or third-party evaluations.
All shared audio consists of live, unscripted recordings chosen by the company.
Open questions remain about how efficiently a speech-only model absorbs facts and new languages compared with text-trained systems, and about how reliably its behavior can be constrained.
For the moment, Dessa stands as a compact demonstration that years of speech research can now fit on consumer hardware and produce conversation that many people find difficult to distinguish from a real person—raising both practical possibilities and the need for careful scrutiny before wider use.
