Block's Dessa voice AI model

Most of what gets called 'voice AI' today - ChatGPT's voice mode, Gemini Live, Sesame's Maya - is a text-based language model wearing a speech costume: audio comes in, gets converted to text, a text-based model reasons over it, and a separate text-to-speech layer bolts a voice onto the output. Dessa, the model Block unveiled on October 23, 2026, skips that pipeline entirely. It was trained from scratch on raw audio alone, with no text transcripts and no pretrained language model underneath it ^[1], and it runs full duplex - it can listen and speak at the same time, the way real conversations actually overlap rather than taking strict turns.

Block's own explanation for why is almost philosophical: babies learn to talk by listening long before they can read a word, and the team wanted to find out whether a model could learn the same way ^[2]. It apparently can, using 'hundreds of times fewer words' of training data than a conventional text-based LLM needs ^[2]. Project lead Ragavan Thurairatnam has pitched the result as nothing less than 'a GPT3 moment for speech' ^[3]- and the compute story backs up the ambition of that claim without the usual price tag: Dessa was trained on a single NVIDIA GB200 NVL72 rack, not a sprawling multi-rack cluster ^[4], and the finished model runs locally, faster than real time, on a MacBook Pro M4 Max ^[5]. The closest prior research comparison, by outside accounts, is Kyutai's Moshi, an open-weight full-duplex speech model with a roughly 7-billion-parameter core, though Moshi and Sesame's CSM both still lean on text somewhere in the pipeline in ways Dessa doesn't ^[6].

The same architecture that makes Dessa sound so startlingly human is also what keeps it out of the tool-calling, agentic-AI stack the rest of the industry has spent the last two years building. Because Dessa never converts speech to text internally ^[1], it has no token stream to attach a system prompt to, no text embeddings to retrieve against, and no obvious hook for calling an API, a calendar, or an MCP tool mid-conversation. A text-based assistant can be instructed, grounded in a knowledge base, or wired into a workflow; Dessa, as described, is closer to a conversational instrument than a conversational agent.

That trade-off is exactly what outside observers have zeroed in on as a real limitation rather than a footnote: an AI that never processes text can't be directed the way enterprises currently expect, and it's an open question whether Block's team can bolt task-following back on without reintroducing the text pipeline they deliberately avoided. It also raises a quieter question the launch materials don't address: a model that learned language purely from conversational audio, without ever ingesting a textbook, an encyclopedia entry, or a line of code, has no obvious way to learn facts or acquire new information the way a text-trained model does - a gap outside analysts have pointed to as unproven rather than solved ^[6].

The clips Block released are genuinely startling - full of overlapping speech, laughter, hesitation, and the kind of cadence that makes a transcript-based assistant sound robotic by comparison. Dorsey himself described the conversations as 'full of emotion, nuance, pauses, and a natural cadence' ^[7]. But the same demo recordings that showcase that realism also show its seams: in one exchange Dessa mistakes its human conversation partner for someone named 'Ronaldo,' and in a guessing game it keeps guessing after it has already been told it won ^[8]. The Verge's Jay Peters, reviewing the same clips, called the back-and-forth 'a little awkward and uncanny' even while crediting the model's realism ^[7].

That split - awe at the texture of the voice, skepticism about the substance of the conversation - shows up in how people are reacting online, too. The loudest praise treats Dessa as proof that full-duplex, text-free speech models have crossed a believability threshold no assistant has hit before; the more skeptical read, especially among people who build or follow competing voice products, is that a curated highlight reel is not the same as a model that holds up outside of cherry-picked demos. Both reactions can be true at once, and Block's choice to publish demo recordings rather than open benchmarks or weights means outsiders currently have no way to settle which read is closer to right.

Block built the realism it's showing off, and it's also the first to say that realism cuts both ways. The company has acknowledged that a voice indistinguishable from a real person's creates new ways for scammers to imitate someone's identity over the phone ^[9]. Critics have been blunter: filmmaker Justine Bateman has asked why AI voices need to sound human at all, arguing that the realism exists 'so you can trick people into lowering their guard' ^[9], and journalist Shiv Aroor has warned that conversational AI this convincing will push phone scam volume to what he called 'pandemic level' ^[9].

Block's response so far has been to gate the model rather than ship it widely: rather than a public release, early access is going specifically to AI safety researchers, with an explicit mandate to probe for risks the company's own team might not have anticipated, including whether Dessa can recognize when a user is in distress instead of just continuing to chat ^[1]. That's a more cautious rollout than most splashy model launches get, but it also means the scam-risk conversation is happening in public, on the strength of demo clips, well before anyone outside a small safety-research cohort can test the actual failure modes for themselves.