Block Trains a Voice Model That Has Never Read a Word

Block's research lab released Dessa this week, a full-duplex speech model trained from scratch on nothing but recorded human conversation. No text, no pretrained models, no ASR stack. Audio goes in, audio comes out, and the whole thing was trained on a single NVIDIA GB200 NVL72 rack.

That is a strange choice in a field where every voice product, from call centers to audiobooks, is built on top of a text pipeline. Block says that pipeline is the problem.

What Block actually built

Dessa listens and talks at the same time. It backchannels, laughs mid-sentence, cuts in when it has something to say, and lets its voice crack when the conversation turns sad. The demos on the release page are live recordings with no steering, and the model picks its own wording, timing and personality.

The lab claims Dessa learned language from hundreds of times fewer words than a text model needs, and that inference runs faster than real time. It runs locally on a MacBook Pro M4 Max or better, putting a conversational model on hardware builders already own. There is no API, no pricing and no public weights yet: Dessa is a research preview with a waitlist, and Block is handing early access to a small group of safety researchers.

Why dropping text is the interesting part

"Interacting with current voice AI can just plain suck," Block writes in the announcement. The complaint is precise: speech gets converted into text, and everything that is not words, accent, breathing, sarcasm, hesitation, tone, is thrown away before the model even starts thinking.

Dessa keeps that information. It hears emotion natively instead of inferring it from punctuation, which Block frames as a safety argument as much as a quality one. A voice agent that cannot hear that someone is in pain will keep cheerfully reading the next line of the script.

What builders can and cannot do with it

Nothing, yet. No SDK, no endpoint, no benchmark table, which is the honest way to read the release: it is a research claim, not a product. The Verge's coverage lands on the same gap, with demos that are impressive and occasionally uncanny.

The direction is the signal. If a speech-only, full-duplex model eventually ships as weights or an API, three things change for voice-agent builders: turn-taking latency drops toward zero, emotion becomes an input rather than a post-processing guess, and per-minute TTS costs can move to local hardware. What does not change is orchestration. A speech model still needs something to decide which tool to call, which CRM to write to, which calendar to check. That layer stays yours.

OpenAI shipped full-duplex with GPT-Live earlier this year, so Block is not alone in chasing natural conversation. It is alone in throwing the text crutch away entirely, and that is what makes Dessa worth watching. If your voice agent still sounds robotic, the text pipeline is the first place to look.