Cyberpunk in reality: in the USA they created an implant that turns brain impulses into text

Scientists from the University of California in The New England Journal of Medicine told that they managed to create a brain implant that is capable of converting neuroimpulses from the brain into text with an accuracy of 93%. The authors of the project admitted that they had been working on this device for ten years, creating various prototypes of their invention, but all previous solutions showed very mediocre results – it was almost impossible to make out the speech. Now they have managed to create a device that will allow people with disabilities to “speak” calmly.

To create the implant, the researchers compiled a full-fledged map of the brain activity of a healthy person, comparing neural impulses with the sounds pronounced by a person, thanks to which words and phrases are formed. After that, the latest development was tested on a 30-year-old patient with complete paralysis, whose brain stem collapsed due to a severe stroke. Before the implant was installed, a person could only spell words using a laser pointer on the visor of his cap, pressing on the on-screen keyboard. Through a surgical operation, a special matrix with electrodes was installed in the area of ​​the brain, which is responsible for the vocal muscles during a conversation.

Further, the matrix installed in the brain read out the neural impulses of the brain, sending them to the computer, where the neural network compared the received information with a pre-compiled map, converting the impulses into words. At the moment, scientists report a fairly high recognition accuracy of words and whole phrases (93%), which is noticeably better than all previous prototypes. But, unfortunately, the speed of converting impulses into words is still very low – 18 words per minute with an average rate of a healthy person of 150-200 words per minute. So there is still room for researchers to grow and develop, but already now the technology looks inspiring.

You may also like