Block's Dessa voice AI laughs, pauses and runs on a MacBook

Block, the company behind Square and Cash App, has shown off a voice AI that sounds unusually human. Called Dessa, it comes from Block's research lab and was announced by co-founder Jack Dorsey on 23 October. TechCrunch reports that the demos are full of laughs, pauses, sighs and even a voice that seems close to tears.

What makes Dessa different

Most voice assistants work in a chain: your speech is turned into text, a language model writes a reply, and a text-to-speech engine reads it out. Block argues that this throws away much of what a voice carries, such as tone, hesitation and mood. Dessa is a speech-to-speech model trained entirely on human speech, with no text in the pipeline. Block says its knowledge comes from the conversational data itself.

It is also full-duplex: it can listen and talk at the same time, so it can be interrupted, react mid-sentence and leave natural gaps. TechCrunch compares it with the GPT-Live model OpenAI released earlier this year. Block says Dessa was trained on a single Nvidia GPU rack.

In one clip posted by Block, a team member plays a guessing game with the model; in another, it says it is tired after work and sounds it. Whether the emotion is steered by hidden tags, as in some rival systems, isn't clear.

Runs on your own computer, but not yet for you

The headline feature for privacy-minded users: Block says Dessa can run locally on a consumer computer such as a MacBook Pro with M4 Max or better, so conversations don't have to go to a server. Our laptop guide for local AI explains why memory and GPU matter for that, and our explainer on open-weight AI models covers what running a model yourself involves.

For now, though, there is no public access. Block has asked AI safety researchers to test the model first, likely because a convincing, emotional synthetic voice is an obvious tool for phone scams and impersonation.

Why Block cares

Square already uses voice AI to take orders for merchants. An on-device model that sounds natural could power many more business and consumer services at a lower cost, TechCrunch notes.

Our take

The demos are impressive, and the local angle is welcome. But a voice that can fake emotion this well raises the stakes for voice cloning and fraud. Block's decision to start with safety researchers is the right one; the real test will be what safeguards come with a public release.