AI Could Aid in Decoding Speech from Brain Activity

Ever since the recent explosion in AI’s popularity people are coming up with more interesting and different ways to utilize this amazing technology. Neuroscience field is no exception. Researchers in Facebook’s parent company Meta state that AI could help thousands of people communicate. People who are not able to do so through speech, gestures, or writing.

The simplified premise of this technology is to scan people’s brains using either magnetoencephalography or electroencephalography, then correlate those readings with a language model to produce words or sentences.

To achieve this goal the team at Meta trained a computational tool to detect words and sentences on 56,000 hours of speech recordings from 53 languages. This includes recognition of speech in a broader level as words or sentences and in a more detailed level – recognizing individual letters and syllables.

Meta’s team then applied this AI language model to various databases that include brain activity readings from 169 volunteers. These volunteers have listened to sentences and words from various books. The AI language model then had the difficult task to sort through vast amounts of data to build an algorithm that recognizes specific letters, words, and sentences just by reading person’s brain waves.

Afterwards, with a help of a computational method that accounts for physical differences among different brains of the test subjects, the team tried to decode what the subjects had been hearing using just three seconds of brain activity data.

AI model then gave more than 1000 suggestions of possible words heard by the test subjects. According to the researchers, the correct answer was in the AI’s top 10 suggestions up to 73 percent of the time.

This seems more than promising for the early stages of development. But why is this research so important? There are existing technologies helping people to communicate by reading and interpreting brain activity. The main selling point is that all existing technologies of this kind require risky brain surgeries to implant electrodes, while this method, although still in its infancy stage, is not surgically intrusive.