Meta's Non-Invasive Brain-to-Text Decoder Reaches 78% Word Accuracy for Its Best Subject

Meta's Brain2Qwerty v2 decodes typed sentences from magnetoencephalography (MEG), the measurement of the magnetic fields produced by neural activity, without any surgery. Trained on 22,000 sentences typed by nine subjects over ten hours each, the system reaches an average word error rate of 39%. Its best participant reached 78% word accuracy, with more than half of sentences decoded at one word error or fewer.

The methodological change matters as much as the accuracy. The first version required knowing keystroke timing in advance, which ruled out real-time use. The v2 pipeline removes that dependency, a step toward decoding intended text as it forms. The underlying approach on natural sentences appears in Nature Neuroscience.

The real gain in capability is limited by the hardware. MEG scanners are large, shielded, and immobile, and the earlier work put electroencephalography, the wearable alternative, at a 67% character error rate, far worse than MEG. This is a laboratory result on memorized sentences, not a wearable product, and it decodes typing intent rather than free inner speech.