Unlock Editor’s Digest Lock for Free
FT editor Roula Khalaf will select your favorite stories in this weekly newsletter.
Neuroscientists strive to give voice to those who cannot speak in a promptly promoted quest to utilize brain waves to restore or strengthen their physical abilities.
Companies such as California university researchers and New York-based precision neuroscience are among those moving forward in generating naturalistic speeches through the combination of brain implants and artificial intelligence.
Investment and attention have long focused on implants that allow severely disabled people to operate computer keyboards, control robotic arms, and regain some use of paralyzed limbs. However, some labs have made progress by focusing on technology that converts thought patterns into speech.
“We’re making great strides, and the big goal is to let a synthetic voice flow from the brain like a chat between two people,” said Edward Chan, a neurosurgeon at the University of California, San Francisco. “The AI ​​algorithms we use are faster and faster, and we are learning alongside all new participants in our research.”
Chang and his colleagues, including the University of California, Berkeley, published a paper in the natural neuroscience last month detailing works with a woman with quadriplegia or paralysis of the limbs and torso that they had been unable to speak for 18 years after suffering a stroke.
She trained a deep learning neural network by trying to quietly say sentences composed using 1,024 different words. The voice of her voice was created by streaming her neural data into a collaborative speech integration and text-deprecating model.
This technique reduced the delay between patient brain signals and results reached 8 to 1 second previously achieved by the group. This is much closer to the 100-200ms time gap of normal audio. The median system decoding speed was 47.5 words per minute, or about a third of a normal conversation.
Thousands of people each year were able to benefit from so-called voice prosthetics. Their cognitive functions remain more or less intact, but suffer from speech loss due to stroke, neurodegenerative disorder ALS, and other brain conditions. If successful, researchers hope to be able to expand their technology to help people who have difficulty speaking due to conditions such as cerebral palsy or autism.
The potential of the voice nervous system is beginning to spark interest among businesses. Precision neuroscience claims that because implant electrodes are more dense, they capture higher resolution brain signals than academic researchers.
The company worked with 31 patients to quickly collect more data and provide a potential route to commercialization.
Accuracy received regulatory clearance on April 17 to keep the sensor embedded for up to 30 days at a time. This will allow scientists to train their systems within a year with what could potentially be “the largest repository of high-resolution neural data present on planet Earth,” said CEO Michael McGer.
The next step is to “make the components smaller, put them in a biocompatible, sealed package and plant them forever in the body.”
Elon Musk’s Neuralink, the most well-known Brain Computer Interface (BCI) company, focuses on allowing people with paralysis to control their computers rather than give a synthetic voice.
A key obstacle to brain-to-voice technology development is the time patients take to learn how to use the system.
An important unanswered question is how much different people have different response patterns in motor cortex (parts of the brain that control spontaneous behavior, including speech). If they remain very similar, previous individual-trained machine learning models can be used for new patients, says Nick Ramsey, a BCI researcher at University Medical Center Utrecht.
It accelerates the process that takes dozens or hundreds of hours of generating sufficient data today by asking participants to view and talk about it.
Ramsey said that all brain-to-voice studies focus on motor cortex, which activates the muscles involved in neurons speaking, and there is no evidence that speech can be generated from other brain regions or deciphering inner thoughts.
“Even if you could, you don’t want people to hear your inner speech,” he added. “There’s a lot of things I don’t say out loud because they don’t benefit me or they might hurt people.”
Sergei Stavisky, co-director of the Institute for Neuropathy Research at the University of California, Davis, said that the development of speech as good as healthy speech could still be “fairly far apart.”
His lab had demonstrated that he could decipher what someone was trying to say with about 98% accuracy, he said. However, the audio output is not instantaneous and does not capture important audio quality such as tone. It was unclear whether recording hardware (electrodes) could be used, he added.
Scientists had to develop a deeper understanding of how the brain encodes speech production and better algorithms to convert neural activity into voice output, Stavisky added.
He said: “In the end, vocal neuropathy must provide a full expressive range of human voices. For example, they can control the pitch and timing accurately and do things like singing.”
