Meta announced Brain2Qwerty v2, a non‑invasive system that converts magnetoencephalography (MEG) recordings into text and reported an average word accuracy of 61% on its test data.

The company said the system uses a helmet‑style MEG scanner to record brain activity while subjects type. Raw neural signals feed an end‑to‑end AI model that the company says also uses large language models to help reconstruct intended sentences when neural signals are noisy.

Meta disclosed that v2 was trained on data from nine volunteers. Each participant wore an MEG device and typed for roughly 10 hours, producing about 22,000 training sentences in total. Meta contrasted the 61% figure with an approximate 8% word‑accuracy level it attributes to prior non‑invasive methods. The company said decoding accuracy continued to improve as training data increased.

Meta also said it will publish the training code for Brain2Qwerty v1 and v2 and that research partners will release the v1 dataset. The work is part of Meta’s Digital Brain Project, which the company says includes a $5 million fund to support open neuroscience datasets.

In a companion paper published in Nature Neuroscience, Meta researchers noted that high‑performance BCIs still largely rely on implanted electrodes, and that non‑invasive approaches face different tradeoffs. Meta framed Brain2Qwerty v2 as a step toward non‑surgical communication tools for people who have lost the ability to speak or type.

The announcement comes amid active work across the field to restore communication with both invasive and non‑invasive systems. Meta’s results are early and based on a small sample; practical use will depend on how accuracy scales with more participants, different devices, and real‑world conditions.

Photo credit: assets.staticimg.com

Tags: MEG, brain-to-text, non-invasive BCI, open dataset, large language models

Topics: Brain–computer interfaces, Wearable neurotech, EEG & neuro-sensing headsets