Meta Achieves High Accuracy in Non-Invasive Brain-to-Text Decoding

Researchers at Meta have made a significant breakthrough in non-invasive brain-computer interface technology, achieving a 61% word accuracy in decoding brain signals into text without the need for surgical implants. This represents a substantial improvement over prior non-surgical methods, which typically yielded around 8% accuracy. The system, known as Brain2Qwerty v2, was trained using data from nine volunteers who typed while their brain activity was monitored. The decoding process involves two main components: an AI model that interprets the raw brain signals and a language model that then reconstructs the intended meaning. One volunteer even reached 78% accuracy, with most of their sentences containing at most one error. Meta has released the training code for both v1 and v2 of this technology, indicating that further improvements in accuracy are expected as more data becomes available, potentially closing the performance gap with invasive brain-computer interfaces.