The algorithm was trained using the neural signals of four women with electrodes implanted in their brains, which were already in place to monitor epileptic seizures. The volunteers repeatedly read sentences aloud while the researchers fed the brain data to the AI to unpick patterns that could be associated with individual words. The average word error rate across a repeated set was as low as 3%.
“A decade after speech was first decoded from human brain signals, accuracy and speed remain far below that of natural speech,” states a paper detailing the research, published in the journal Nature Neuroscience. “We trained a recurrent neural network to encode each sentence-length sequence of neural activity into an abstract representation, and then to decode this representation, word by word, into an English sentence,” the report states.
The system is, however, still a long way off being able to understand regular speech. “People could become telepathic to some degree, able to converse not only without speaking but without words,” the report stated.