Dessa, Block's AI Voice That Laughs and Cries Like a Human

Block, the payment company founded by Jack Dorsey, has just unveiled something that goes beyond a common voice assistant. The model is called Dessa, has emerged from the company's research lab, and does something that most competitors cannot: it sounds genuinely human. With laughter, pauses, sighs, and even the tone of someone about to cry, the technology raises both admiration and serious concerns about fraud.

The novelty arrives at a time when the artificial intelligence market is in an intense race to make voice interactions more natural. But while most models still convert speech to text before processing the response, Dessa tries something different. And this significantly changes the equation.

How Dessa Works and Why It’s Different

Most AI voice systems follow a predictable path: you speak, the model transcribes to text, processes the information, generates a text response, and then converts it to audio. It’s a multi-step chain, and each step discards emotional nuances. The tone of impatience in your voice, the hesitation before a question, the subtle sarcasm—this all gets lost in transcription.

Dessa operates as a full-duplex model, meaning it speaks and listens simultaneously, much like OpenAI's GPT-Live launched this year. The central difference that Block claims is that emotion is not added through tags or external commands. It is embedded in the model itself, trained entirely with data from real human speech.

In demonstrations published by the research team, the model adopted different voices, cadences, and emotional states. In one conversation, Dessa said it was tired after a day of work, and the tone effectively sounded exhausted. In another, when talking about someone who made it feel comfortable, the voice became choked, almost breaking.

Runs on a Laptop, Not in a Datacenter

Perhaps the most relevant data for the tech market is the infrastructure required. Dessa was trained on a single rack of NVIDIA GPUs and can run locally on a consumer computer, such as a MacBook with an M4 Max chip or higher. This is remarkable.

Most advanced voice models rely on heavy cloud servers, which creates latency and high costs. A model that runs on-device opens possibilities for real-time applications, without a constant internet connection, and with drastically lower operational costs. For a company like Block, which already operates in the payments and small business solutions universe, the practical implication is direct: high-quality automated customer service running on the merchant's own device.

Square, Block's arm focused on small businesses, already uses voice AI to manage orders. With Dessa, the company could elevate this experience to a level where the customer doesn’t even realize they are speaking to a machine.

The Problem of Excessive Realism

And here lies the dilemma that inevitably accompanies this type of advancement. If an AI sounds indistinguishable from a human, the possibilities for use in scams grow in proportion. Imagine receiving a call with the emotional voice of someone asking for help, and that voice is entirely synthetic.

Block itself acknowledges the risk. The company has not released public access to Dessa and has published a call for security researchers to test the model. The stated goal is to identify scenarios in which a voice agent could deceive people on the other end of the line. It’s a prudent measure, but it highlights how much the technology has advanced into delicate territory.

The audio deepfake market already incurs billion-dollar losses. According to the U.S. Federal Trade Commission, voice imitation fraud caused losses of over $2.7 billion in 2023. A model like Dessa, which adds realistic emotional layers, makes detection even more difficult.

What This Means for the AI Landscape

Block's decision to enter the voice model game reveals something about the market's direction. Fintech companies do not want to rely on external suppliers like OpenAI, ElevenLabs, or Google for critical customer service functionalities. Developing proprietary models that run locally is a cost-reduction strategy and, at the same time, a competitive differentiation.

Block is not the first to tread this path. But the approach of training exclusively with human speech data, without going through the intermediary of text, could represent a turning point for the industry. If the model delivers on its promises at scale, the integration between voice AI and financial services enters a new phase.

For now, Dessa remains closed. There is no timeline for public launch. But the fact that a model with this level of expressiveness runs on a laptop already tells a story about where technology is heading. And about the risks that come along with it.

For the financial sector, the equation is clear: whoever masters the voice interface with AI will have a brutal advantage in service, costs, and user experience. The remaining question is whether regulation will be able to keep pace.