Block's Dessa Learned to Talk From Sound Alone: a Full-Duplex Speech Model With No Text and No LLM, Trained on a Single GPU Rack

Block, the payments company behind Square and Cash App, has stepped into foundation-model research with an unusual bet. On Thursday its small in-house AI lab unveiled Dessa, a full-duplex speech-to-speech model that the company says was trained from scratch on nothing but recorded human conversation — no text corpus, no transcription pipeline, and no pretrained language model underneath.

The pitch targets the biggest weakness of today's voice assistants. Most of them convert speech to text, reason over that text, then synthesize a spoken reply, a round trip that strips out timing, emotion and the small sounds that make conversation feel human. Dessa skips the intermediate step entirely: audio goes in, audio comes out, and the model listens and talks at the same time, able to interject, backchannel with a "yeah" or "got it," and react to laughter, sighs and tone the way a person on a phone call would.

The training setup, per Block, was a single NVIDIA GB200 NVL72 system — a large budget by single-company standards but tiny next to frontier text-model clusters — and the company claims the model learned language from "hundreds of times fewer words" than large language models see. Block says inference runs faster than real time and that Dessa runs locally on a MacBook Pro with an M4 Max chip or better. No parameter count, dataset size or benchmark results were published, so all performance claims remain the company's own and are not independently verified.

In place of benchmarks, Block released recordings it describes as live and unsteered. In them, Dessa laughs mid-sentence, raises its volume when asked, finishes a speaker's sentence when they blank on a movie title, and gasps at a surprise anniversary announcement. The company was candid about the awkward moments too: in one widely shared demo the model mistook its conversation partner for someone named "Ronaldo," an error Dessa lead Ragavan Thurairatnam posted without trying to hide it.

The safety framing is the serious part. Block argues that voice AI is "still nearly deaf to the emotion in our voices" — its example is a person in serious pain while a voice assistant cheerfully carries on — and that an agent which can hear distress, urgency or frustration, and pause or stop, is a meaningfully different proposition. Early access is therefore limited to a small group of AI safety researchers, who must apply with an institutional email, "to find the risks we haven't anticipated." A public waitlist is open to everyone else.

The lab reports directly to Block cofounder and chairman Jack Dorsey and is led by Thurairatnam, who co-founded Dessa, the Toronto deep-learning startup Block acquired in 2020. Six researchers are listed on the project, and the team says it has built generative speech models for more than eight years. Thurairatnam called the release "a GPT-3 moment for speech" on X, while Dorsey confirmed in replies that Dessa is "speech to speech, no LLM" and that it runs locally.

Whether a speech-only model can actually learn facts and reasoning as efficiently as text-trained ones is one of the open research questions Block itself lists, alongside multilingual ability and in-context learning. For now Dessa is a research preview, not a product — no weights, no pricing, no release date. But it is a notable signal that a consumer fintech is shipping original foundation-model research, and that the text-pipeline design of voice AI has a credible challenger. If the approach scales, the first place it will show up is anywhere a machine has to hold a real conversation — support lines, agents, and companions — without the flat, turn-taking cadence users have learned to tolerate.