Malhar Ujawane

Inspiration

Inspiration·

Block's Dessa speech model

Block's Dessa is a full-duplex speech model trained only on human conversation audio. What it gains by skipping text, and what it can't do yet.

Block's Dessa speech model

Dessa: the Conversation Model trained only on speech (via Hacker News)

Dessa is a speech-to-speech model from a small research lab inside Block, announced on October 23, 2026. Block says it was trained from scratch on recordings of human conversation only: no text, no pretrained models. It’s full-duplex, so it listens and talks at the same time, and it learned when to cut in and when to drop a “yeah” or “got it” from the conversations it heard.

The line from the page that explains why they bothered:

Run speech through text and all of that is gone, which is why most voice AI feels stilted, slow and, worst of all, annoying.

“All of that” is the emotion, sarcasm, laughter and sighs in the sentence before it. I wrote about how text becomes integer token IDs in What are tokens in AI?, and a transcript is a lossy step before that one even starts: the words survive, the tone of voice doesn’t. Most voice agents I’d sketch today are a pipeline of speech-to-text, then an LLM, then text-to-speech. Each hop adds latency and throws away anything that isn’t a word.

A few specifics from Block’s page. Inference runs faster than real time, and the model runs locally on a MacBook Pro with an M4 Max or better. They trained it on a single NVIDIA GB200 NVL72, and they say it learned language from “hundreds of times fewer words” than an LLM needs. The lab reports to Jack Dorsey and was founded by former members of Dessa, the Toronto deep learning startup.

What you can’t do yet is try it. The page has recorded conversations, no API and no weights, and early access is going to a small group of AI safety researchers. Phrases like “the most authentic full-duplex speech model ever built” are Block’s claim, and right now the only evidence is their own clips.

The question I’d ask before building anything on it: how does an audio-only model call a tool? Agents work because the model can emit a structured tool call, like a JSON blob that says “look up this order”, and a model with no text in training has no obvious way to produce one. A voice agent that sounds human but can’t check a calendar is a demo. If Block solves that part, a model this size running on a laptop gets interesting for on-device voice features.

Visit Block's Dessa speech modelvia Hacker News (295 points on 2026-10-24)

Get new posts in your inbox

© 2026 Malhar Ujawane. Built with Astro.