The technology that allows you to transcribe your work meetings could help people with paralysis to speak again.
Researchers at UC Berkeley and UC San Francisco have used a generative AI to reduce the delay between the moment when a person suffering from severe paralysis tries to speak and when the computer device plays sound. Their work helped a woman named Ann, who underwent a stroke of cerebral trunk in 2005 at 30 years, to communicate in real time. Ann spoke with a voice that sounded like hers because the model was formed on records of her before her stroke.
The deployment of the generation in different ways has enabled researchers to make improvements in neuroprosthhesis that could have taken much more time, said Cheol Jun Cho, a Ph.D. UC Berkeley. Student in electrical engineering and computer science and co-author of the study, which appeared in March in Nature neuroscience.
This is an example of how generative AI tools – using the same underlying technology that fuels chatbots like OpenAi Chatpt and Claude or Anthropic Transcriptions in Google Meet – help medical and scientific researchers to solve problems that could have taken much more time to solve, said Cho. AI experts and donors have highlighted the use of medicine technology as a field with a huge advantage, whether to design new drugs or provide better tests and diagnostics.
“AI is accelerating progress,” said Cho. “Sometimes we imagined that the chronology would be a decade or two. Now, this rhythm is like three years.”
The technology that has helped Ann is proof of concept, said Cho, but it shows a path to tools that could be more plug and play in the future.
Japier the speech
The problem with existing neuroprosthesis is latency. There is a gap between the moment when the person begins to try to speak and when a sentence is really generated and heard. Cho said the previous technology meant that Ann had to wait for a sentence to be completed before starting the next one.
Ann, seen during the first research study in 2023, was able to communicate through computers that read the signals that his brain tried to send to the muscles that control the word.
“The major breakthrough here is that she doesn’t need to wait until she finishes the sentence,” he said. “Now we can really disseminate the decoding procedure each time she intends to speak.”
The prosthesis includes a range of electrodes located on the surface of its brain and connected via a cable to a computers’ bank. It decodes the control signals that Ann’s brain sends to the muscles that control the word. After Ann chose the words she intends to say, an AI reads these signals from the engine cortex and gives them life.
To form the model, the team has been trying to talk about the sentences displayed on a prompt on a screen. They then used data on this activity to map the signals of the engine cortex, using the AI generation to fill the gaps.
Cho said the team hopes that the breakthrough will lead to scalable and more accessible devices.
“We are still in continuous efforts to make it more precise and with weaker latency,” he said. “We are trying to build something that can be more plug and play.”
Use AI to go from thought to speech
Cho said the team had used the AI generation in different ways. One was to reproduce the voice before Ann’s injury. They used recordings before his injury to form a model that could produce the sound of his voice.
“She was very excited when she heard her own voice again,” said Cho.
https://www.youtube.com/watch?
The big change was in real -time transcription. Cho compared it to the tools that transcribe presentations or meetings when they occur.
The work built on a 2023 Study which used AI tools to help Ann to communicate. This work still had a significant delay between the moment when Ann tried to speak and when the words were produced. This research has considerably reduced this delay, and Ann told the team that it seemed more natural.
“She pointed out that she felt embodied, that it was her own speech,” said Cho.
