Meta Reports Typing Decoded From Non-Invasive Brain Scans at 61 Percent Word Accuracy

Meta's research group has described Brain2Qwerty version 2, a pipeline that reconstructs typed sentences from magnetoencephalography (MEG) signals recorded while a person types, with no implant and no surgery. The system reaches 61 percent word accuracy on average, with the best participant at 78 percent and more than half of sentences decoded at one word error or fewer, according to reporting on the release.

The improvement over prior non-invasive work is large. Meta states that earlier non-invasive methods reached only about 8 percent word accuracy. The model was trained on roughly 22,000 sentences from nine volunteers, each of whom wore a 306-sensor MEG device for around ten hours.

The constraints are as important as the accuracy. The MEG scanner is expensive, does not move, and requires a magnetically shielded room, which rules out everyday assistive use for now. The results also come from healthy volunteers typing in a controlled setting, not from patients with the speech or motor impairments that a clinical device would need to serve. What is established is a method advance in decoding language from non-invasive signals, not a deployable product.