From Brain Waves to Words: Brain2Qwerty Offers a New Path to Communication Without Surgery

Meta just dropped the sequel to a research project it started last year, and the upgrade is bigger than a typical version bump. Brain2Qwerty v2 takes brain activity recorded through a non-invasive helmet and turns it into typed sentences in real time, no electrodes implanted in anyone's skull required. That distinction matters more than it might sound: the invasive route, using techniques like stereotactic electroencephalography or electrocorticography, has already proven that AI can turn brain signals into communication for people who've lost the ability to speak or type. But surgery doesn't scale. You can't offer a brain implant to millions of people who need it.

So Meta's team trained this new version on roughly 22,000 sentences gathered from nine volunteers, each of whom sat through 10 hours of recording while wearing a magnetoencephalography device and actually typing. Rather than building a hand-tuned pipeline that hunts for specific neural signatures, the researchers went end-to-end: raw brain signals in, deep learning model figures out the rest. They also fine-tuned large language models on the neural data itself, which lets the system lean on semantic context to fill in gaps when the brain signal is noisy and ambiguous. Engineers even used AI agents to help hunt for pipeline optimizations, though humans still picked the final training setup by hand.

The numbers are the real headline here. Word accuracy hit 61%, a sharp jump from the 8% seen in other non-invasive approaches. For the best-performing participant, accuracy climbed to 78%, meaning more than half their sentences came out with a single word error or none at all. That's a meaningful leap toward the kind of performance previously reserved for surgical methods, and it comes without anyone going under the knife.

What might matter even more long-term is a pattern Meta noticed: decoding accuracy scales log-linearly with the amount of data fed into the system. In plain terms, more recordings from more people could keep closing the gap with invasive techniques, no fundamental breakthrough required, just more data. That's a much easier problem to throw resources at than solving new neuroscience from scratch.

Meta is releasing the full training code for both v1 and v2, and its research partner, the Basque Center on Cognition, Brain, and Language, is putting out the v1 dataset alongside it. The work sits under Meta's broader push toward open brain-foundation models, including tools like Tribev2 for perception encoding and NeuralSet for processing brain data at scale, backed by a $5 million fund meant to encourage more open datasets in this space. The stated goal is straightforward: build this stuff in public so neuroscience moves faster than it would if everyone worked behind closed doors.