Block's new voice AI Dessa was trained entirely on human speech and runs locally on a MacBook

Block, the payments company led by Jack Dorsey, has unveiled Dessa, a voice AI model that its research lab trained entirely on recordings of human speech. There was no text and no pretrained language model in the training mix. Block calls it a full-duplex, speech-to-speech model, which means it listens and talks at the same time instead of waiting for you to finish a sentence.

Most voice assistants today turn your voice into text, run that text through a language model, then turn the reply back into audio. Block argues that this pipeline throws away much of what people actually communicate: emotion, sarcasm, laughter, breathing, pauses and emphasis. Dessa works on audio directly. In the demos Block published, it interrupts naturally, drops quick acknowledgements such as "yeah" and "got it", laughs in the middle of a sentence and changes its tone to match the mood. Asked on X which language model sits behind the voice, Dorsey replied: "speech to speech. no llm."

The hardware numbers are the part builders will notice. Block says it trained Dessa from scratch on a single Nvidia GB200 NVL72 system, and that the finished model runs locally on a MacBook Pro with an M4 Max chip or better, with no internet connection needed. Block has not disclosed the model's size, benchmark results or a release date.

For now Dessa is a research preview, not a product. Block is giving early access to a small group of AI safety researchers and asking them to look for risks its own team may have missed. Reaction online split two ways. Some users said they had never heard anything artificial sound so human, while fraud watchers warned that a voice this convincing could make phone scams harder to spot.

Why it matters for developers: a natural, low latency voice that runs on a laptop removes the cloud round trip that makes today's voice agents feel slow and costly to run. If the approach holds up, on device voice could become a real option for privacy sensitive apps, customer support and accessibility tools. Every claim so far comes from Block's own announcement and demos, though, not independent tests.

What to watch next: whether Block publishes weights, a paper or an API, and how the safety researchers' findings shape a wider release. If you build voice features, test interruption handling, laughter and emotional range yourself before trusting the demos.

Building with AI? See this week's hot AI repos and how to use each one.