Block Dessa Voice Model Runs Locally and Sounds Eerily Human

Block, the payments company co-founded by Jack Dorsey, has unveiled Dessa, a full-duplex speech model that the company says was trained entirely on human conversation and can run locally on consumer hardware. The announcement, made on October 23, 2026, positions voice as the next major interface for AI agents, one that no longer needs to route every word through the cloud.

According to Block, Dessa is a speech-to-speech model: audio goes in and audio comes out, with no text stage in between. The company claims that design preserves the emotional information that conventional voice pipelines discard, including tone, hesitation, laughter, and sighs. In demonstration recordings published by Block researchers, Dessa shifts cadence and emotion to match what it is saying, sounding tired while describing a long day at work and emotional while discussing a moment of comfort.

Block co-founder Jack Dorsey described the experience in a post on X, saying that talking with Dessa feels like talking with a person, full of emotion, nuance, pauses, and a natural cadence. Ragavan Thurairatnam, who leads the team behind the model, called the release a milestone for speech and said Dessa can run on consumer GPUs, describing it as a defining moment for conversational voice technology.

Block Dessa Voice Model Runs Locally on Consumer Hardware

The local angle is what sets Dessa apart from most voice assistants. According to reporting by TechCrunch, Dessa runs locally on a MacBook Pro M4 Max or better, faster than real time, after being trained on a single NVIDIA GPU rack. That is a modest training footprint compared with the large clusters used for frontier text models, and it points to a future in which capable voice agents live on the device rather than in a data center.

Running locally carries three practical benefits. Conversations do not have to leave the device, which matters for privacy-sensitive settings such as health care, finance, and the home. Responses arrive without a network round trip, removing the awkward pauses that make many voice assistants feel mechanical. And once the model is running, there is no per-minute API bill, which changes the economics for businesses that want voice agents in every store, clinic, or vehicle.

The trade-off is breadth. Text-trained models can draw on vast written knowledge and call external tools, while Dessa draws its knowledge from conversational data. As reported by Unite.AI, Block presents Dessa as a research model rather than a general-purpose assistant, and the most likely near-term path is a hybrid: a Dessa-style voice front end paired with a capable agent behind it. That combination would matter for the wider agent ecosystem, where voice is increasingly the interface through which people direct agents to book, buy, schedule, and troubleshoot.

Related coverage of the agent ecosystem includes recent reporting on Anthropic merging Claude chat and Cowork for enterprise users and on Gemini launchpads bringing AI agents to federal agencies, both of which show agents moving from chat windows into the workflows where work actually happens.

Humanlike Voice Draws Praise and Raises Scam Concerns

The reaction to Dessa has split in two directions. Fans praised the natural quality of the voice. As reported by PYMNTS, one listener said the voice sounded more human than anything artificial heard before, while Binance founder Changpeng Zhao pointed to education, where natural voice interaction could help children learn through conversation.

Critics raised a different concern: a voice that sounds convincingly human can also deceive. Filmmaker Justine Bateman argued that a human-sounding AI voice could trick people into lowering their guard, according to PYMNTS, while journalist Shiv Aroor warned that scam calls could reach a new scale. The worry is concrete for voice agents. A system that can interrupt, backchannel, laugh, and sound tired is far harder to distinguish from a person on a phone call than a flat robotic voice, and that realism raises the stakes for caller verification, consent, and disclosure.

Block appears to be taking that risk seriously. There is no public access to Dessa at launch. Instead, the company has opened a waitlist and invited AI safety researchers to test the model first, a step that will likely focus on cases where a voice agent could mislead the person on the other end of a conversation. That staged rollout mirrors a broader shift across the industry toward treating safety evaluation as a precondition for deploying autonomous systems, not an afterthought.

What Dessa Means for Voice Agents

For builders of voice agents, Dessa signals three shifts. First, the pipeline is collapsing: speech-to-speech models remove the text round trip that made older assistants sound stilted. Second, the hardware bar is falling: a model that runs on a laptop-grade machine can be deployed in places where cloud connectivity is unreliable, expensive, or simply unwanted. Third, the safety bar is rising: the more human a voice sounds, the more important it becomes to prove who, or what, is speaking.

Practical uses are already easy to picture. A warehouse supervisor could ask about inventory while walking the floor, a field technician could diagnose equipment without pausing to type, and a finance manager could ask which invoices are becoming high risk and have suppliers contacted, all through conversation rather than screens. PYMNTS has previously argued that many business processes already function like conversations, and voice agents that compress several corporate interactions into one exchange could reshape how that work gets done.

Whether Dessa itself becomes a product or remains a research milestone, the direction is clear. Voice agents are moving on-device, becoming more emotionally expressive, and getting harder to tell apart from people. The companies that win this next phase will be the ones that pair that realism with safeguards strong enough to keep trust intact, because a voice agent that people cannot trust is one they will simply hang up on.

Sources: Block announcement via PYMNTS and TechCrunch coverage of Dessa, with additional reporting from Unite.AI on the model design and rollout.