Meta's Non-Invasive Brain-to-Text Decoder Reaches 61 Percent Word Accuracy

Meta's Fundamental AI Research (FAIR) lab published Brain2Qwerty, a pipeline that decodes typed sentences from magnetoencephalography (MEG) recordings, with the v2 system reaching an average 61 percent word accuracy and 78 percent for its best participant. The underlying method appears in Nature Neuroscience, and per MarkTechPost the training code and dataset are public.

The technical claim is a large increase for surgery-free decoding. Prior non-invasive approaches were near 8 percent word accuracy, far below the electrode implants used in clinical brain-computer interface (BCI) work. Meta trained on roughly 22,000 sentences from nine volunteers, each wearing an MEG helmet for about ten hours, and replaced hand-tuned signal-processing stages with an end-to-end deep network coupled to a language-model prior that corrects likely character sequences.

The caveats are physical, not statistical. MEG requires a magnetically shielded room and a helmet-scale sensor array, so this is a laboratory result, not a wearable device. The system also decodes typing, which produces motor signals, rather than imagined speech. What is verified is a reproducible accuracy gain on a public dataset. What is not yet shown is whether the approach transfers to portable sensors or to users who cannot move.