August 23, 2023 | 5:32 p.m
A woman who did not utter a word for years after a paralyzing stroke has regained the ability to speak through artificial intelligence.
The groundbreaking method used an array of 253 electrodes, which were placed in the brain of Ann Johnson, 48, and then connected to a bank of computers via a small port connection mounted on her head.
Electrodes, which cover the part of the brain where speech is processed, intercept his brain signals and send them to a computer, creating a brown-haired avatar representing Johnson.
The on-screen avatar — which Johnson chose himself — is then able to “speak” what he’s thinking using a copy of his voice recorded during a 15-minute toast at his wedding.
The avatar also blinks its eyes and uses facial expressions such as smiles, lips and raised eyebrows, making it appear more lifelike.
“We’re just trying to get people to recover,” said Dr. Edward Chang, chairman of neurological surgery at the University of California, San Francisco. told the New York Times.
Johnson — a high school math teacher who was also active as a volleyball and basketball coach in Saskatchewan — had been married for two years and had two children when a stroke left him paralyzed.
“It’s bad not being able to hug and kiss my kids, but that was my reality,” Johnson said. “The real nail in the coffin was being told I couldn’t have any more children.”
After years of rehabilitation, he gradually regained some movement and facial expression, but Johnson was unable to speak and had to be tube-fed until swallowing therapy allowed him to eat finely chopped or soft foods.
“My daughter and I love cupcakes,” Johnson said.
The team at UCSF, along with colleagues at the University of California, Berkeley, is the first time that speech or facial expressions have been synthesized from brain signals.
To train the AI system, Johnson had to silently “repeat” from a 1,024-word vocabulary until the computer recognized the pattern of brain activity associated with each word.
Instead of whole words, the AI program was taught to recognize phonemes, the units of speech that make up spoken words. “Hello,” for example, has four sounds: “HH,” “AH,” “L” and “OW.”
By detecting 39 sounds, the AI program could decode Johnson’s brain signals into full sounds at a rate of about 80 words per minute — about half the rate of a normal person-to-person conversation.
Sean Metzger, who developed the decoder at UC Berkeley and UCSF’s joint bioengineering program, told Southwest News Service that the program’s “accuracy, speed and vocabulary are critical.
“This is what gives a user the possibility to communicate over time as quickly as we do and to have a much more natural and natural conversation.”
The team is now working on a wireless version, which means the user does not have to be physically connected to the computer with wires or cables.
Chang has worked on brain-computer interfaces for more than a decade and hopes the team’s innovations will lead to a system that enables speech to be generated from brain signals in the near future.
“Our goal is to restore a full, embodied way of communication, which is really the most natural way for us to talk to others,” Chang told SWNS.
“This advance brings us much closer to making this a practical solution for patients,” Chang added.