Reading a sentence in the brain, without an implant

Meta AI's Brain2Qwerty v2 decodes raw MEG brain signals into text with up to 78% accuracy using a fine-tuned LLM, marking a leap in non-invasive decoding.

Meta AI shares Brain2Qwerty v2, a new milestone in its research on non-invasive brain-to-text decoding.

The first version, published in Nature Neuroscience, already showed that a model could reconstruct typed sentences from EEG or MEG recordings, with a clear advantage for magnetoencephalography. V2 takes the system further: approximately 22,000 sentences were recorded from nine volunteers, each equipped with a MEG helmet for ten hours, then used to train a pipeline capable of converting raw brain signals into coherent sentences. Meta reports an average word-level accuracy of 61%, and up to 78% for the best participant, with approximately half of the sentences decoded with at most one error.

The difference also lies in the architecture: Brain2Qwerty v2 combines character, word, and sentence representations, with a fine-tuned LLM to leverage semantic context. The Brain2Qwerty v1 and v2 code is released, while the BCBL makes the v1 dataset available. An important caveat remains: these results still concern healthy volunteers, in a laboratory MEG setting. The transition to patients deprived of communication remains the decisive step.