Block releases Dessa, a voice model that laughs, pauses and almost cries

Block, the financial technology company co-founded by Jack Dorsey, has released a new voice model called Dessa from its research lab — and it laughs, pauses, sighs and even sounds as if it is about to cry. The company said the model is trained fully on human speech and runs on a single NVIDIA GPU. That combination — expressive emotion and small-scale local hardware — puts Dessa in direct competition with some of the largest voice AI products on the market, several of which depend on cloud infrastructure and text-driven pipelines.

What makes Dessa different from existing voice AI?

Dessa is a full-duplex model, meaning it can talk and listen at the same time, in the manner of the GPT-Live model OpenAI released earlier this year. Most speech models on the market today are paired with a large language model (an AI system trained on text that handles knowledge retrieval and reasoning). Block said Dessa takes a different route: its knowledge stems entirely from the conversational data it was trained on.

The company's central technical claim concerns the pipeline. Block said current voice models can sound unnatural and annoying because text sits in the pipeline that drives speech generation. In practice, products such as ElevenLabs v4 and Fish Audio Drama 3 use voice tags — explicit labels inserted into the input — to steer emotion in generated speech. Block claims emotion is built into Dessa itself, not bolted on through tags. The company has not clarified whether some form of tagging nevertheless steers that emotion behind the scenes, leaving an open question about the model's actual mechanism.

Block's language in its blog post is blunt. The company said:

"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."

The passage makes two arguments at once. The first is experiential: transcribing speech to text discards prosody — the pauses, laughs, tremors and timing that carry meaning. The second is safety-oriented: a system that cannot hear how a user feels may respond inappropriately, and Block frames that limitation as potentially dangerous rather than merely suboptimal.

What did the demos show?

Block's researchers posted multiple example conversations in which Dessa's agent displayed different voices, cadences and emotions. In one conversation, the bot said it was tired from work — and at that point in the exchange, it audibly sounded tired, according to the company's demonstration.

In another example, the bot spoke about the person on the other end making it comfortable, while sounding emotional and appearing to almost break down mid-sentence. A third demonstration showed a Block team member playing a game with the voice assistant.

The demos matter because they show expressiveness as a continuous property of the model's speech rather than a set of discrete, tagged emotional states. A voice that conveys fatigue while discussing fatigue is doing something text-driven systems approximate only through explicit instructions. Whether Dessa sustains that quality across long, unstructured conversations is not demonstrated in the source material.

Where can Dessa run — and who can use it?

Block said Dessa can run locally on consumer-grade hardware, naming the MacBook M4 Max or better as the baseline. Local execution carries two implications. First, latency: on-device speech processing avoids round trips to cloud servers. Second, cost: a model that runs on one machine does not incur per-inference cloud fees, a meaningful consideration for any company deploying voice agents at scale.

There is no public access to Dessa at the moment. Block has instead put out a call for safety researchers to test the model — an invitation the company frames around probing how the agent behaves, likely including cases where a voice agent could deceive people on the other end of a call. A voice that sounds tired, laughs naturally or nearly cries is also, by construction, a voice that can be mistaken for a human's. That dual-use character explains why safety testing precedes public release.

Why is a payments company building voice models?

Block is in the business of finance and sells its solutions to small businesses. Its Square division already uses voice AI to handle ordering — an existing deployment where conversational quality directly affects transaction completion. A proprietary voice model that runs on-device would let Block power more business and consumer experiences at a lower cost than licensing or operating cloud-based alternatives.

The strategic logic runs deeper than ordering systems. Voice AI is becoming a primary interface for customer-facing services across retail, food service and payments. Companies that own their speech stack control latency, cost, expressiveness and, critically, the data flowing through customer conversations. For a company whose customers are small merchants, a voice agent that handles calls, orders and support without cloud dependency is an infrastructure play, not a research curiosity.

How does Dessa fit the wider voice AI market?

The competitive context is worth stating plainly. OpenAI shipped GPT-Live earlier this year with full-duplex conversation. ElevenLabs released v4, and Fish Audio shipped Drama 3, both using voice tags to drive emotional output. Block is entering that field from an unusual position: it is not an AI lab or a model vendor, and its flagship model is designed to run on a laptop rather than a data center.

The single-GPU claim is the most concrete differentiator. If a fully expressive, full-duplex voice model genuinely runs on one NVIDIA GPU or a MacBook M4 Max, the economics of voice AI deployment shift — for Block's merchants first, and potentially for anyone the company eventually licenses the technology to. The claim remains Block's own, backed by demonstrations rather than an independent benchmark, and the company has not published training details, dataset composition or evaluation results.

What happens next?

Two things will determine whether Dessa is a research demonstration or a product. First, the safety review: Block has invited external researchers to test the model, and the scope of that testing — particularly around deception scenarios — will shape any release timeline. Second, integration: Square already handles ordering with voice AI, and an on-device model could extend that footprint across Block's small-business customer base at lower cost. Until the model reaches outside hands, the strongest verifiable facts remain Block's own: full-duplex speech, emotion without confirmed tags, and a footprint small enough for a consumer laptop.