Block's Dessa voice model runs on one GPU and mimics human emotion

Block, the fintech company co-founded by Jack Dorsey, has released a voice model that runs locally on a single NVIDIA GPU and produces laughs, pauses, and cries that sound unsettlingly human. The company said the model, called Dessa, is trained entirely on human speech — with no text pipeline in between.

The claim matters because most voice AI on the market today works differently. Models such as ElevenLabs v4 or Fish Audio Drama 3 convert speech to text and use voice tags to drive emotion in speech generation. Block argues that approach throws away signal.

What does Dessa actually do differently?

Dessa is a full-duplex model, meaning it can talk and listen simultaneously — a capability OpenAI brought to market earlier this year with its GPT-Live model. Speech models are typically paired with a large language model to retrieve knowledge. Block said Dessa's knowledge comes only from the conversational data it was trained on.

The company's critique of competitors is blunt. In a blog post, Block wrote: "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. We think that's ridiculous, leads to a worse experience, and is potentially dangerous."

That is a strong statement from a company that has published no independent benchmarks. Whether Dessa's emotion is genuinely learned or steered by some form of internal tagging remains an open question — the company does not say.

What did the demos show?

Block's researchers posted multiple example conversations in which the agent displayed different voices, cadences, and emotions. In one exchange, the bot said it was tired from work — and its voice audibly sounded tired at that moment. In another, the bot discussed how the person made it feel comfortable while sounding emotional, nearly breaking down. A third demo showed a Block team member playing a game with the voice assistant.

Demos are demos, of course. They are selected by the vendor, and no external party has verified how the model behaves in unscripted settings.

Why is Block not releasing it publicly?

There is no public access to Dessa at this time. Instead, Block has put out a call for safety researchers to test the model — a step the company frames as due diligence, and which is almost certainly aimed at cases where a voice agent could deceive people on the other end of a call.

That caution cuts both ways. It signals awareness of the deception risk inherent in hyper-realistic voice AI, but it also means the company's central claims — on-device performance, built-in emotion, full-duplex fluency — remain self-attested.

What does this mean for Block's business?

Block is fundamentally a finance company that sells solutions to small businesses. Its Square unit already uses voice AI to handle ordering. A proprietary voice model that runs on-device could power far more business and consumer experiences at lower cost, since it does not depend on cloud inference or per-call API fees.

The company said Dessa can run locally on a consumer-grade machine, such as a MacBook M4 Max or better. That hardware bar means mass-market phones are, for now, out of reach — but business terminals and modern laptops are not.

The takeaway is plain: if Dessa's claims hold up under outside testing, Block gains a voice capability it can embed across Square and its consumer products without paying a model provider for every interaction. Until safety researchers get their hands on it, every promise here rests on Block's own word.