Block's research lab opened a research preview of Dessa this week: a speech-to-speech model trained from scratch on recorded human conversation — no text, no pretrained language model, no transcription step anywhere in the loop.
That last detail is the point. Almost every voice assistant shipping today is a chain. Speech recognition turns your voice into text, a language model writes a reply, a text-to-speech engine reads it aloud. Prosody — the pause, the half-laugh, the flat tone that says someone is tired — is either discarded at the first hop or painted back on afterwards with style tags. Block's argument, published alongside the demos, is blunt: "Interacting with current voice AI can just plain suck. 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."
Dessa is full-duplex, meaning it listens while it talks, so it can be interrupted and can interrupt. In the clips Block released, the model laughs mid-sentence, pauses as if thinking, and in one recording sounds close to tears while describing being made comfortable. Block says the conversations were recorded live with nobody steering the responses. Jack Dorsey's own summary on X: "Talking with it feels like talking with a person… full of emotion, nuance, pauses, and a natural cadence."
What Block has disclosed:
- Trained entirely on human speech, from scratch, with no text data and no pretrained model. - Trained on a single Nvidia GB200 NVL72 system, and compact enough to run locally on a MacBook Pro with an M4 Max or better. - Full-duplex audio in, audio out. No separate speech-recognition or text-to-speech component to bolt on.
What it has not disclosed: model size, any benchmark result, or a release date. Early access is going to a small group of AI safety researchers, with a waitlist behind them. And Dessa's knowledge comes from the conversational data it was trained on, per Block — it is a conversation model, not an oracle. Do not expect it to do your research.
The name is not new. Square acquired the Toronto voice-AI startup Dessa in 2020, and the research lab that built this model — which reports directly to Dorsey — includes staff from that company. The commercial fit is obvious: Square already uses voice AI to handle phone ordering, and an on-device model that needs no per-minute API bill changes the economics of answering the telephone. It also, inevitably, changes who can afford a voice that sounds like a person.
Why this matters beyond a demo: for four years text has been the compression layer for AI, and it is lossy in exactly the place human trust is decided. A model that hears hesitation can respond to hesitation. That is the promise. It is also the risk, and Block is not pretending otherwise — hence safety researchers before users. Critics made the point within hours of the demos going up. Filmmaker Justine Bateman asked why an AI voice needs to sound human at all: "So you can trick people into lowering their guard." Journalist Shiv Aroor put it shorter: "Scam calls about to attain pandemic level." Voice cloning and phone fraud are already cheap; a model that improvises a plausible emotional reaction is a different order of problem for the person on the other end of the line.
The counter-argument is that the same fidelity is what makes a voice interface usable for someone who cannot type, or for a child learning through conversation — the use case Binance founder Changpeng Zhao raised publicly, pointing at his own education project. Block adds that speech-native models could eventually detect distress or pain that a transcript flattens into nothing.
None of that is testable yet. There are no independent evaluations, no published numbers, no access for anyone outside the preview. Treat the demos as a company's own evidence about its own product. Until outside results exist, Dessa is a compelling demonstration and a credible thesis: the text bottleneck in voice AI is a design choice, not a law of nature.
