Jack Dorsey, Twitter co-founder and chairman of Block, introduced Dessa this week, a voice AI model that operates entirely without text at any stage of its pipeline. Unlike every existing voice system — which converts speech to text, processes it through an LLM, and generates new speech from text — Dessa works purely in audio: sound in, sound out, trained entirely on real human conversations.
The announcement came on October 23 through a post by Dorsey on X and the official block.xyz/dessa page, with video demonstrations of conversations ranging from amusing to emotionally unsettling. The model was developed by Block's internal research lab, formed by ex-members of Dessa Original (the Toronto startup founded in 2019, which Square acquired and renamed), and operates in real time on consumer hardware — specifically, a MacBook Pro with an M4 Max chip or higher.
Breaking with the current paradigm
Every voice system that exists today goes through an intermediate step of speech-to-text conversion. This eliminates fundamental nuances: intonation, hesitation, dramatic pauses, the tone of a contained laugh. The Block team itself describes this process as "flattening what we say into text and throwing away an incredibly large amount of information."
"The most powerful AIs in the world can't hear how you feel," the company wrote in an explanatory post. "We think that's ridiculous, leads to a worse experience, and is potentially dangerous. Imagine being in serious pain while a voice assistant has no idea and just keeps asking what you want."
Dessa's approach is analogous to how humans learn language: by listening. "Babies learn to talk without ever having seen a written word," the project's site says. "We wanted to see if a model could do the same."
The task is, in fact, more difficult than it appears. Speech data never repeats: saying the same sentence twice produces different audio files, even pronounced identically. The space of possibilities is essentially infinite. Text data, by contrast, is identical regardless of who types it. That's why Block claims Dessa learned language with hundreds of times fewer words than large language models, but the trade-off is a radically different architectural approach.
What Dessa does
The model is full-duplex: it speaks and listens simultaneously. This allows natural interruptions, overlapping speech, and the ability to "backchannel" — making sounds like "uh-huh," "got it," "ah" — to signal attention during conversation.
In the published demonstrations, Dessa shows an impressive range of expressions: laughing in the middle of a sentence when something is funny, speaking louder when asked, a frustrated sigh, even tears audible in the voice when expressing gratitude. One demonstration shows the model conversing with GPT-Live-1 (from OpenAI), with both voices AI-generated — a conversation between two voice models.
Inference runs faster than real time, eliminating the artificial pause that characterizes current assistants. This means a response begins before the user finishes speaking, exactly like a natural human conversation.
Why this matters: cost, safety, and the future
Block is, above all, a payments and fintech company. Square already uses voice AI to process restaurant orders. With a model that runs locally — no cloud calls — inference costs plummet dramatically. The company stated it trained Dessa on a single NVIDIA GB200 NVL72 system, but inference runs on consumer hardware, opening possibilities that cloud-based models simply can't reach.
But there's a safety dimension that worries experts. If a voice model is indistinguishable from a human voice, phone phishing becomes much harder to detect. CoinDesk pointed out this concern specifically: "it's the ultimate scammer's weapon." Block itself acknowledges it's opening uncharted territory and is inviting safety researchers to test the model before public release.
Ragavan Thurairatnam, the lab's lead, called the release "a GPT-3 moment for audio." The analogy is revealing: just as GPT-3 showed that scale qualitatively changes model behavior, Dessa suggests that scaling speech data may produce conversational capabilities that hybrid (text + voice) systems could never achieve.
What remains open
The project page lists open research questions: how efficiently do these models learn language and facts compared to text-based models? How do they behave with new languages? What kind of in-context learning can they perform? No answers yet.
Dessa has no disclosed model size, public benchmarks, or release date. Early access is limited to a group of safety researchers. But what has been published is enough to generate a sense that the voice AI field — stagnant in the text-to-speech paradigm for years — may be entering a second phase.
The technology that learns by listening is, for now, only a prototype. But if real conversation depends on micro-nuances that text inevitably strips away, the path to truly conversational intelligence may lie in pure audio — and Block has just taken the first step.
If AI learns language by listening the way humans do, when does it begin to behave as if it were actually present in the conversation — not just responding, but truly listening?
Sources: TechCrunch, The Verge, CoinDesk
✓ Independent sources cross-checked and verified before publishing
