For the next five months, machine learning gurus can try to best predict the speech of a brain-computer interface (BCI) user who lost the ability to speak due to a neurodegenerative disease. Competitors will design algorithms that predict words from the patient’s brain data. The individual or team whose algorithm makes the fewest errors between predicted sentences and actual attempted sentences will win a US $5,000 prize.
The competition, called Brain-to-text ‘25, is the second-annual public, open-source brain-to-text competition hosted by a research lab part of the BrainGate consortium, which has been pioneering BCI clinical trials since the early 2000s. This year, the competition is being run by the University of California Davis’s Neuroprosthetics Lab. (A group from Stanford University hosted the first competition using brain data from a different BCI user.)
For two years, the UC Davis research team has collected brain data from a 46-year-old man, Casey Harrell, whose speech is unintelligible except to his regular caregivers. Once the speech BCI was trained on Harrell’s brain data, it could decode what he was trying to say over 97 percent of the time and could instantly synthesize his own voice, as previously reported by IEEE Spectrum.
Decoding Speech from Brain Data
Parsing words from brain data is a two-step process: The algorithm must first predict speech sounds, called phonemes, from neural data. Then it must predict words from the phonemes. Competitors will train their algorithms on the brain data corresponding to 10,948 sentences with accompanying transcripts of what Harrell was attempting to say.
Then comes the real test: The algorithms must predict the words in 1,450 sentences from brain data withheld from the training data. The difference between the final set of predicted words and the words that Harrell attempted to say is called the word error rate—the lower the word error rate, the better the speech BCI works, overall.
Researchers reported a 6.70 percent word error rate, which they hope the public can beat. The goal of the competition is to attract machine learning experts who may not realize how valuable their skills are to speech BCIs, says Nick Card, a postdoctoral researcher at UC Davis leading both the clinical trial and the competition.
“We could sit on this data and hide it internally and make more discoveries with it over time,” says Card. “But if the goal is to help make this technology mature faster to help the people who need to benefit from this technology right now, then we want to share it and we want people to help us solve this problem.”
The public invite into the research world is “an awesome development” that is “long overdue” in the BCI space, said Konrad Kording, a professor at the University of Pennsylvania who researches the brain using machine learning, and who is not involved in the research or competition.
This year, Card and his fellow researchers have raised the bar by lowering the starting word error rate with their own high-performing algorithm. The first brain-to-text competition in 2024 began with the Stanford University group posting an error rate of 11.06 percent and finished with the competition winner achieving 5.77 percent. Also new this year are cash prizes for lowest error rates and the most innovative approach, provided by BCI company Blackrock Neurotech, whose electrodes and recording hardware have been used by BrainGate clinical trials since 2006.
Ethical Concerns in BCI Data Sharing
BCIs have long served as a bridge between neuroscience, medicine, and machine learning. And while machine learning has a tradition of open-source research, medical research is bound by patient confidentiality.
The main concern with public brain data is that the patient will be identified, says bioethicist Veljko Dubljević, a professor of both philosophy and science, technology, and society at North Carolina State University.
That concern is moot in this case because Harrell went public in August 2024, roughly five years after he began losing muscle tone because of amyotrophic lateral sclerosis (also known as Lou Gehrig’s disease). In 2023, neurosurgeons at UC Davis implanted four electrode arrays with a total of 256 electrodes into the top layers of his brain. Harrell used his speech BCI in an interview with the New England Journal of Medicine last year to explain how the disease feels like being in a “slow-motion car crash.” Harrell said at the time that “it was very painful to lose the ability to communicate, especially with my daughter.”
The speech BCI was trained on data collected while Harrell conducted in-lab experiments and while he spoke casually with family and friends. But competitors of Brain-to-text ‘25 will not see any “personal use” data recorded while Harrell spoke casually and extemporaneously, Card says.
While this is a “good precaution,” Dubljević says, he wonders if Harrell realizes what it means to have someone’s sensitive medical data in the public domain for years. The “noise” of today’s BCIs could be decoded into meaningful personal information in 50 years, for instance, in a way similar to how blood donated in 1955 can now also reveal details about a person’s DNA. (DNA profiling wasn’t established until the 1980s.) Dubljević recommends limiting the data storage to five years.
Speech BCIs decode the intended movements of a person’s jaw and mouth muscles, in the same way a BCI for an arm or hand prosthesis decodes intended movements. But speech BCIs feel more personal than BCIs that control a hand prostheses, Dubljević says. Speech is closer to “the innermost sanctum of a person,” he says. “There’s quite a lot of fear about mind reading, right?”
“As a researcher who wants to see science technology deployed for the public good, I want the technology not to be hyped up” in order to avoid a backlash, Dubljević says.
Cash Prizes for Innovative BCI Solutions
The two lowest word error rates come with $5,000 and $3,000 cash prizes, respectively, and the most innovative approach will win $1,000.
The last category is meant to encourage out-of-the-box ideas with great potential, if given more data or more time. Stacking 10 multiples of the same algorithm is a common way to force a more accurate overall performance, but it costs 10 times as much computational power and, “quite frankly, it’s not a very creative solution, right?” Card says.
The innovative category is likely to attract the usual crowd of academic and industry BCI scientists, who enjoy finding creative solutions, Kording says.
But the top slots will likely go to coders with no background in BCIs and who sport a “street fighting” style of machine learning, as Kording calls it. These “street fighters” focus on speed over ingenuity. In practice, the best BCI algorithms, Kording said, are “usually not really driving from a deep knowledge of how brains work. They’re driving from a deep understanding of how machine learning works.”
That said, both the traditional BCIs and new entrants are important parts of the science engineering ecosystem, Kording says. With the corners full, the competition is slated to be an exciting battle.
