Block’s Dessa Voice Model Can Listen and Speak at the Same Time on a MacBook M4 Max

Block — the payments and technology company founded by Jack Dorsey — says its research team has built a voice model that operates in full-duplex mode. That means it can listen and generate speech at the same time, much like a human in a real conversation.

The company says Dessa will run locally on a MacBook Pro fitted with Apple’s M4 Max chip or better. No internet connection required for processing. The audio stays on the device.

Block says the model runs faster than real time on compatible hardware. That claim has not been independently verified.

Trained on Human Speech, Not Text

Most large language models learn from written text scraped from the internet. Block says Dessa took a different path — it was trained on recorded human conversation.

The result, according to reported demonstrations, is a model that produces natural-sounding pauses, laughter, interruptions, and backchannel responses — the small verbal cues like “mm-hmm” or “yeah” that people use to show they’re listening. Whether it genuinely sounds indistinguishable from a human is a subjective judgement, and no independent testing has confirmed that claim.

Block reportedly trained Dessa using a single NVIDIA GB200 NVL72 system — a rack-scale AI supercomputer. That detail has not been independently confirmed from an official technical specification.

What Full-Duplex Actually Means

Most voice AI systems work in half-duplex. They wait for you to finish speaking, process what you said, then reply. It’s the same model as a walkie-talkie — one direction at a time.

Full-duplex changes that. The model can keep generating speech while still taking in new audio. If you interrupt, it can react. That’s closer to how humans actually talk.

But building a system that handles both directions simultaneously without falling apart is technically hard. Block’s claim that Dessa manages this on a laptop chip — rather than on data-centre hardware — is what makes the announcement stand out. The company has not published parameter counts, architecture details, or independent benchmark results, so the full picture remains unclear.

Privacy and Safety

Running AI locally has real privacy implications. When a voice model processes audio in the cloud, that audio travels across a network and sits on someone else’s server. Local processing keeps it on your own machine — in principle, at least. Privacy still depends on the operating system, the software around the model, and how data is stored.

Block has invited safety researchers to test Dessa before any wider release. That’s relevant because voice models carry specific risks. A convincing synthetic voice can be used to impersonate people, generate misleading audio, or enable fraud. External safety testing is one way to surface those problems before they reach the public.

The model is not publicly available. There is no download, no API, and no broadly accessible version for developers or consumers to test. Block describes this as a research preview.

Amir Shevat, who leads developer relations at Block, has spoken publicly about the company’s broader ambitions in AI tooling, though no specific quote from a named Block researcher about Dessa’s technical design has been confirmed in available reporting.

What We Don’t Know Yet

There is quite a lot Block hasn’t said. The model’s parameter count is unknown. Licensing terms haven’t been published. There’s no confirmed timeline for a public release, a commercial product, or an open-weights version that developers could run and test themselves.

That limits what anyone outside Block can say about how well Dessa actually performs. Demonstrations controlled by the company that built the model are a starting point, not a verdict.

The technology sector has seen a run of voice AI announcements over the past year — from OpenAI’s real-time voice features in ChatGPT to Google’s Gemini Live. Block is entering a field where the competition is well-funded and moving fast. Running on a consumer MacBook rather than a data centre would be a meaningful practical advantage if the performance claims hold up under independent scrutiny.

What This Means for Kent Residents

For most people in Kent, Dessa isn’t something you can download or use right now — there’s no public release and no consumer product. But the broader direction matters: voice AI that runs on a laptop rather than in the cloud could eventually mean faster, more private voice assistants that work without a strong internet connection, which is relevant in parts of the county where connectivity remains patchy. Developers and researchers at Kent’s universities or technology businesses may want to watch Block’s next steps, especially around whether model weights or an API become available.