Speech neuroprosthesis
The brain implant helped to reproduce the silent speech of a paralytic by neural activity
Anastasia Kuznetsova, N+1
American doctors have installed a brain implant for a paralyzed man, which reads neural signals in speech motor zones, and according to this activity, the computer composes whole phrases. So far, his vocabulary is small – 50 words – but in the future they plan to improve the technology and teach the system more words. The study was published in The New England Journal of Medicine (Moses et al., Neuroprosthesis for Decoding Speech in a Paralyzed Person with Anarthria).
In people after a stroke, brain centers are often affected, which are responsible not only for limb movements, but also for speech. If the loss of the ability to speak is combined with paralysis of the upper and lower extremities, then it is difficult for the patient to use auxiliary devices that need to be controlled by hands. We have already told you that American scientists have developed a prototype of an invasive decoder device that allows synthesizing speech by brain activity during the movement of the organs of the speech apparatus. The device was first tested on volunteers who uttered phrases aloud, and the computer was trained to understand the pattern of activation of neurons. In paralyzed people, such training is more difficult to carry out if they have lost the ability to speak. It is also unclear to what extent such patients, after decades of paralysis, retained the ability to activate neural networks responsible for speech. Previously, attempts were made to read individual sounds from a paralyzed person, but they have not yet tried to distinguish words, much less phrases.
A team of American doctors led by Edward F. Chang from the University of California has set a more ambitious goal than their predecessors. Scientists decided to teach a computer to read individual words from a paralyzed 36-year-old man who had a massive stroke at the age of 20. They installed an implant on the patient above the surface of the motor areas of the brain of the left hemisphere responsible for speech. In this case, the neural network was initially trained on a patient who did not speak, and used only brain activity for this. First, the model learned to read the activity of neurons, and then predict the possible continuation of a phrase from a vocabulary of 50 words. Thus, the synthesizer composed separate short phrases. The training process lasted several months.
After the training period, the computer understood the patient in 74.4 percent of cases, which is still far from ideal. The implant can distinguish speech activity at a rate of 15.2 words per minute. In the tests, the patient was able to answer simple questions about how he was doing and whether he wanted to drink.
Figure from the article by Moses et al.
The researchers plan to attract more participants to clinical trials of the speech implant, expand the vocabulary of the system and increase the speed of speech recognition. Further research should also improve the accuracy of reading words.
Portal "Eternal youth" http://vechnayamolodost.ru