Meta has created a “brain decoder” that uses an AI model to decode brain activity and convert it into speech. The tool can be used for imaginary communication between users who say words and phrases in their head. The company reported the research on its corporate blog.

Meta technology differs from similar developments in that it relies on non-invasive methods of recording brain activity. The company used electroencephalography (EEG) and magnetoencephalography (MEG), which only use external sensors. Data from EEG and MEG are less accurate than information from sensors placed in the brain, so much larger amounts of raw data had to be processed for greater efficiency.

The AI ​​recorded how the participants' brains responded to the audiobooks and individual phrases, and then extracted the necessary words from the text and created a "dictionary" that did the reverse process - converting thoughts into speech. The researchers achieved an accuracy of 73% when using a string of 793 words most frequently used in daily life.

In the future, AI will be able to more accurately identify words and phrases using a larger source dictionary, according to the researchers. The development should help not only millions of people who cannot speak or write, but also the study of the human brain.