Block, the payments company led by Jack Dorsey, has entered the voice AI race with an unusually bold claim. On Thursday, the company's in-house research lab unveiled Dessa, a speech-to-speech model that Block says was trained from scratch on nothing but human conversation — no text, no transcripts, and no pretrained foundation models underneath.
"As far as we know, Dessa is the world's first AI you can have a realistic live conversation with that was trained only from human speech," Block says on the model's official page, which describes the release as "a research preview of a full-duplex speech model." The company is publishing live demo recordings, opening a public waitlist, and granting early access to a small group of AI safety researchers. For more context on this story, see our ongoing latest AI developments.
What Dessa Actually Is
Dessa is what researchers call a full-duplex speech-to-speech model: audio in, audio out, with no text layer in between. That last part is the differentiator. Virtually every mainstream voice assistant — including the voice modes attached to the largest chatbots — converts your speech to text, reasons over it, and speaks a response back. Block argues that pipeline throws away most of what makes human conversation human.
"Interacting with current voice AI can just plain suck," the company's announcement states. "It turns your voice into text and throws away an incredible amount of information text alone can't carry. The most powerful AIs in the world can't hear how you feel. We think that's ridiculous, leads to a worse experience, and is potentially dangerous."
Because Dessa never passes through text, Block says it natively picks up on emotion, sarcasm, laughter, a frustrated sigh, or a voice cracking with sadness — the paralinguistic signals that text-based systems typically discard.
Full-Duplex Means It Talks Like a Person, Not a Turn-Taking Bot
The second defining trait is full-duplex operation. Dessa listens and talks at the same time, the way people do, rather than waiting for a detected end-of-utterance before responding. According to Block, the model learned from real conversations when to cut in, when to backchannel to show it is following along, and when to stay quiet.
The demos on the company's page illustrate behaviors nobody explicitly programmed. In one clip, Dessa laughs mid-sentence the moment something becomes funny — "the way people do," Block notes, contrasting this with conventional voice assistants that "wait for you to finish, then laugh on cue." In another, a user asks the model to speak up and it gets louder. "Nobody taught it that," the company says. In a third, the person blanks on a movie title and Dessa finishes the sentence for them.
Block is also careful to flag the provenance of one particularly eye-catching demo, in which Dessa holds a live conversation with another AI system. "In this demo, Dessa is talking with GPT-Live-1, a third-party AI model. Both voices are AI-generated," the page states, noting that GPT-Live-1's developer is not affiliated with Block and has not reviewed or endorsed Dessa.
Why Block Skipped Text Entirely
Training a model with no text at all is a deliberate bet. Block's position is that speech is not a lossy wrapper around language — it is the primary medium, and a model that grows up on raw sound may develop capabilities and failure modes that text-trained systems do not have.
The company itself frames the scientific upside as open questions rather than settled results: how efficiently do speech-only models learn language and facts compared to text-trained systems? How efficiently can they learn new languages? What kinds of in-context learning can they perform? "Dessa advances the frontier of deep learning models and creates a different category with entirely new and interesting behaviors for researchers to discover," the announcement reads.
There is also a pragmatic claim buried in the pitch. Block says Dessa's inference runs faster than real time, so replies come without the awkward pauses that plague turn-based voice assistants, and — according to coverage by Unite.AI of the launch — that the model runs locally on consumer hardware as capable as a MacBook Pro with an M4 Max chip or better. If accurate, a locally running full-duplex voice model would be notable at a moment when most competitive voice AI remains cloud-bound.
A "GPT-3 Moment for Speech"
Ragavan Thurairatnam, who leads Block's AI research lab, described Dessa as the first AI model that can hold a realistic conversation after training entirely on speech, and called the release "a GPT3 moment for speech," according to Unite.AI's report on the announcement. The comparison is pointed: GPT-3's 2020 debut did not immediately produce useful products, but it reframed what language models could do and kicked off the current investment cycle.
The Verge, which covered the launch, described the result as a model that "talks to you like an awkward new friend" — an evocative summary of demo clips in which Dessa gasps "oh my god" under its breath at a surprise, or responds to personal news with what sounds like teary gratitude. Whether those mannerisms reflect genuine understanding or sophisticated pattern reproduction is precisely the kind of question Block says it wants the research community to probe.
The Safety Angle Comes First — For Once
Notably, Block is not shipping Dessa straight to consumers. The release is a research preview: recordings are public, a waitlist is open, but hands-on early access is being limited to a small group of AI safety researchers. Given ongoing concern about voice-cloning scams and emotionally manipulative AI companions, a speech model that can express — and presumably elicit — emotion is exactly the kind of system safety teams want to evaluate before mass deployment.
Block says more recordings will be shared on X in the coming days. The company has not announced pricing, API access, or integration plans for its Square, Cash App, or Bitkey products.
The Open Questions
Dessa's launch lands in a crowded field. Every major AI lab now ships a voice mode, and startups have been pushing real-time speech models for two years. What distinguishes Dessa is its training recipe — pure speech, zero text — and the claim that emergent conversational behavior like mid-sentence laughter arose without explicit instruction.
Skeptics will want independent evaluation. Can a text-free model actually learn facts reliably, or does it hallucinate more freely without the grounding of written language? Can it scale to new languages without transcripts? Block's own page invites exactly these questions, which is either confidence or a candid acknowledgment that nobody — including Block — knows the answers yet.
For now, Dessa is a research artifact with a waitlist. But if speech-only training proves competitive with text-grounded pipelines, Block may have surfaced one of the more interesting contrarian bets in modern AI: that the fastest route to natural machine conversation is to build machines that learn the way we did — by listening.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
