Knowledge
4:30 p.m. April 1, 2020To the short link
–
Neuroscientists from the University of California in San Francisco have managed to generate text from the brain signals of test subjects. The results of the study were published in the journal Nature Neuroscience.
–
According to the EurekAlert portal, the research team led by Joseph Makin and Edward Chang worked with epilepsy patients. They had up to 250 electrodes implanted in their cortex to observe their disease.
The subjects had to read and repeat sentences. The electrical impulses of their nerve cells were measured and an electrocorticogram was made.
Artificial intelligence analyzes brain waves
A program then analyzed the audio recordings made during the experiment and sifted through the brain signals for patterns. These suggested certain language elements, such as consonants and vowels.
The researchers then fed the cortical representations explored from repetitions of the sentences into the second program. It should read a text out of it.
Word error rate as in speech recognition programs
The error rate of the two-part AI was three percent. Many mistakes had no connection to the original text, neither a phonetic nor a semantic.
Nevertheless, the scientists pointed out that the word error rate is comparable to that of the typical speech recognition programs, the portal said.
The approach could eventually help people who are no longer able to articulate themselves.
mo / mt
– .
