Meta introduced the revolutionary Brain2Qwerty v2 system, which is capable of recovering text from brain activity records without the need for surgery. The system’s word-level accuracy has reached 61%, more than seven times higher than previous non-invasive methods, which barely reached 8%.

Some 22,000 sentences were used to train the model. Nine healthy volunteers participated in the experiment, each spending 10 hours on a magnetic encephalography scanner, tapping the phrases they heard. Instead of outdated manual neural event detection schemes, Meta applied deep learning to decode data directly from raw brain signals.

The system works with a continuous brain activity record and restores sentences in their entirety, using a pretrained language model to account for the context of meaning. The best participant in the experiment achieved an impressive 78% accuracy, with more than half of his proposals decoded with one error or no errors at all.

Despite the successes, development is still far from clinical or domestic use. Brain2Qwerty v2 was only tested on healthy people who were physically typing. For patients who have lost speech or ability to move, this approach requires a separate check. In addition, the system does not work in real-time mode - it processes the entire sentence, so the user does not see every word immediately after its "input" by thought.

The project is part of the $5 million Digital Brain Project initiative to create open neuroscience datasets. Recall that back in September 2022, Meta published a study on the "brainstorming decoder" that transforms thoughts into speech.