Skip to content
Komposite News
Markets · Technology · Capital
Thursday, September 10, 2026Global EditionGot a tip?
Voice AI

Block trains a voice model on sound alone

Block, the Square and Cash App owner, has unveiled Dessa, a speech-to-speech model it says was trained only on recorded human conversation, with no text. It says the model runs on a high-end laptop. It is not publicly available.

Tom Kessler · House byline
October 25, 20263 min read
The earpiece and coiled cord of a red telephone handset, close upTrintage via Unsplash
The earpiece and coiled cord of a red telephone handset, in a file photograph.Photograph via Unsplash

Block, the payments company behind Square and Cash App, has built a voice model called Dessa that it says learned to hold a conversation from recorded human speech alone, with no text and no pretrained model, according to the company's announcement. Jack Dorsey, Block's co-founder, promoted it on X on Friday, TechCrunch reported.

"We trained Dessa from scratch on nothing but human conversation," Block's announcement says. "No text, no pretrained models." The company describes Dessa as a speech-to-speech model, meaning audio goes in and audio comes out without being turned into text along the way, and as full-duplex, meaning it can listen and talk at the same time. That lets it cut in, laugh mid-sentence and drop in short acknowledgements such as "yeah!" while the other person is speaking, Block says. It calls Dessa "the most authentic full-duplex speech model ever built" and, "as far as we know", the first AI trained only on human speech that can hold a realistic live conversation. Neither claim has been independently verified.

The pitch is aimed squarely at the dominant design, in which a voice assistant transcribes speech, sends text to a large language model and reads the answer back. "Interacting with current voice AI can just plain suck," the announcement says, arguing that the text step "throws away an incredible amount of information" such as tone, sarcasm and distress. Block frames that partly as a safety issue: a system deaf to emotion, it says, could carry on regardless while a user is in serious pain.

Block's efficiency claims are its own. It says Dessa learned language from "hundreds of times fewer words" than large language models need, without giving a figure, and that it was trained on "only one NVIDIA GB200 NVL72", a single rack-scale system of Nvidia graphics processors. Inference runs faster than real time and the model runs locally on "a MacBook Pro M4 Max or better", the company says. TechCrunch noted that Block says the model's knowledge comes from its conversational training data rather than from a separate language model, and that it is not clear whether the emotion in its voice is steered by tags of the kind other speech models use.

The evidence offered so far is a set of recordings. Block's page hosts several live conversations between people and Dessa, plus one between Dessa and GPT-Live-1, which TechCrunch identifies as OpenAI's voice model; Block notes that GPT-Live-1's developer is not affiliated with it and has not reviewed or endorsed Dessa. Neither Block nor TechCrunch gave benchmark scores, pricing or licensing terms, and there is no public access to the model, according to TechCrunch.

Instead, Block is offering early access to "a small group of AI safety researchers" to look for risks it has not anticipated. TechCrunch said the call was likely to test cases in which a voice agent could trick the person on the other end. The announcement is signed by Ragavan Thurairatnam, the Dessa lead. Block says the "(very) small" lab was founded by former members of Dessa, the Toronto deep-learning start-up, and reports directly to Mr Dorsey.

Why it matters: Block's businesses run on merchants and consumers, and voice is an obvious interface for both. TechCrunch noted that Square already uses voice AI to handle ordering, though Block has not said how or whether Dessa will be used in its products. The broader claim, that a capable conversational model can be trained without text on a single rack, would bear on the economics of voice AI if it holds up under outside testing. Komposite's recent coverage of smaller models built to run on local hardware includes Nvidia's open Nemotron and Bolt releases, and it has covered other attempts to build models that skip text output, such as TypeSafe AI's Jev.

MORE ON THIS STORY