Winner of Chen Institute and Science Prize uses AI to rebuild speech from brain signals
When Sergey Stavisky first started thinking about brain-computer interfaces (BCI) as an undergraduate at Brown University, he was motivated by three factors. “I liked building things,” he recalled, “and I wanted to do something medical. But I was also fascinated by the mind.”
That combination would lead Stavisky into a field that is now rapidly redefining what it means to lose, and potentially regain, a voice.
Today, Stavisky is an associate professor of neurological surgery at the University of California, Davis, and a leading figure in the development of AI-powered speech neuroprostheses. His work, recognized this year by the Chen Institute and Science Prize for Al Accelerated Research, sits at the intersection of neuroscience, clinical care and machine learning. But at its core is a simple goal: restoring the ability to speak to people who have lost it.
That goal becomes vivid in the story of one participant in his team’s research, a man living with amyotrophic lateral sclerosis (ALS) who could no longer speak intelligibly.
Through an implantable device and a suite of AI models trained on his brain activity that Stavisky and his team designed, the man is now able to generate fluent sentences – first as text, then as synthetic speech modeled on his own pre-ALS voice. In moments of daily use, he has produced millions of words.
Dealing with data overload
The science behind that achievement depends on a reality that has transformed neuroscience over the past decade: data overload. “Brain signals are really complicated,” Stavisky explained. Where researchers once recorded from single neurons, modern systems can capture signals from hundreds of neurons at a time. But the behaviors they are trying to decode, like speech, are among the most complex human actions, and traditional statistical methods for processing data, Stavisky said, simply break under such complexity.
He and his colleagues needed a way to process massive amounts of neural data very quickly and flexibly. “AI turned out to be uniquely powerful for that,” he said.
In Stavisky’s system, one model decodes brain activity into phonemes, the basic sound units of language. Another model, drawing on large language modeling approaches, converts those phonemes into words and sentences. In an alternative straight-to-voice approach, deep learning systems reconstruct speech sounds, producing a synthetic voice in real time.
The result is a system that can translate intention into speech with fidelity, generating audible sounds with delays as short as 30 milliseconds. This is fast enough to approximate natural conversation.
In a study published in Nature Medicine in June, Stavisky and colleagues describe how their BCI helped the participant with ALS to maintain rich interpersonal communication with his family and friends at home, independently control his personal computer, and sustain full-time employment.
“Stavisky developed an AI-based speech neuroprosthesis with immediate and transformative practical impact,” said Yury V. Suleymanov, senior editor at Science. “It restored communication for a paralyzed patient with amyotrophic lateral sclerosis with over 99% word accuracy, enabling the patient to express 2.7 million words over two years using only brain signals. His team achieved real-time voice synthesis, allowing the patient to modulate intonation and even sing.”
From movement to speech
Stavisky said a moment early in his career, while working on BCIs for movement, led him to pivot to focus on BCIs for speech. He had noticed something consistent across patients: restoring the ability to move a cursor or robotic arm was valuable, but restoring communication was always more urgent. “Communication was always the number one priority,” he said.
That realization, combined with emerging advances in machine learning and intracortical recording technology, led him to change mid-career from motor prosthetics to speech. At that time, speech decoding from brain signals was widely considered one of the hardest problems in neuroprosthetics. But progress in AI was accelerating at exactly the right moment. Even consumer dictation systems were beginning to reach usable performance levels around 2018, he said.
Looking ahead, Stavisky said the long-term goal is a “high-fidelity surrogate voice”—a system so natural that if someone were speaking on the phone, “you couldn’t tell it wasn’t their natural voice.” The future will likely involve devices that are smaller, fully implanted, and less visible than today’s research systems. It will also require moving from laboratory prototypes to widely available clinical tools.
Already, the field is expanding. Companies are beginning to enter clinical trials for speech BCIs, and academic labs are exploring whether similar approaches could help people with stroke-induced aphasia, cerebral palsy or other language disorders. The implications, Stavisky suggested, could extend far beyond paralysis.
"Ten years ago, Tianqiao and I founded the Chen Institute to pursue a fundamental question: how does the brain give rise to intelligence?,” said Chrissy Luo, Chen Institute cofounder. “We could not have imagined then that AI would change not just how we study the brain, but what we could learn from it. Dr. Stavisky's research has done something once considered nearly impossible: decode brain signals directly into speech, giving patients back the ability to communicate in their own voice. This prize was created for exactly this kind of work, and we are proud to celebrate his achievement alongside our partners at AAAS and Science. We remain committed to championing the researchers who are redefining what science can achieve."
Finalists
Finalists for the prize include Zhiling Zheng, for his essay, “Reprogramming Synthesis: General-Purpose AI Agents in the Materials Chemistry Laboratory.” The second finalist for the prize is Nicholas C. Jacobson, for his essay, “Generative AI to Scale Precision Evidence-Based Psychotherapy.”
Can brain-computer interface training improve your ability to catch mistakes?
Study shows that participants can learn to modulate brain electrical activity to improve perception of minor visuo-motor errors
The brain uses visual cues to coordinate muscle movement. When the motor commands and sensory feedback are out of alignment, visuo-motor errors occur. Rapid perception of these errors allows for correction, which is important in all aspects of life—from preventing falls in the aging to enabling precision in surgery. A new study, published by Wiley in Advanced Science, showed that training with feedback from brain electrical activity, called brain-computer interface training, improves detection of subtle visuo-motor errors.
Quantified using electroencephalogram (EEG) tests, the brain emits characteristic electrical signature, called the error-related potential (ErrP), when individuals recognize an error committed by themselves or others. One component of the ErrP, a positive deflection known as the error positivity (Pe), specifically emerges when an individual becomes consciously aware of the error. Researchers hypothesized that Pe can be modified through learning to enhance perception of visuo-motor errors.
To determine whether feedback on the brain’s electrical activity can improve perceptual learning, researchers compared their brain-computer interface training with traditional behavioral training. Participants completed a task in which they used a joystick to move a cursor towards a target in a straight line. In random trials, the cursor trajectory was altered with different rotation magnitudes to introduce a visuo-motor error. The behavioral training group recorded whether they observed a rotation in each trial and subsequently received feedback on their response. After completing the same task, the brain-computer interface training group saw whether their EEG registered an ErrP as feedback. Participants in both groups completed training every day for five consecutive days.
The researchers found that the amplitude of the Pe increased when the participant perceived a rotation in the trial and, over the five days of training, Pe amplitude increased overall as participants' error perception improved. Behavioral training improved the perception of visuo-motor errors for larger rotations, but not smaller rotations. In contrast, brain-interface training resulted in accelerated learning and improved perception of smaller visuo-motor errors. EEG revealed contributions from the parts of the brain that control decision-making and visuospatial processing.
These findings suggest that brain-computer interface training is more effective than conventional behavioral training at improving the perception of small visuo-motor errors. Safer than pharmacological strategies for improving perceptual learning, future applications of this intervention include strengthening cognitive function in neuropsychiatric patients and facilitating dynamic responses in motorsport drivers.
“This approach targets the neural signature of error awareness itself, not just behavior. By decoding the Pe component in real time and feeding it back to participants, we help the brain amplify its own marker of conscious error detection—something conventional training can't do once errors get too subtle to notice. That lets us drive learning gains for exactly the small errors that behavioral training alone couldn't touch,” said senior author José del R. Millán, PhD, of the University of Texas at Austin in the United States.
Additional information
NOTE: The information contained in this release is protected by copyright. Please include journal attribution in all coverage. For more information or to obtain a PDF of any study, please contact: Sara Henning-Stout, newsroom@wiley.com
Full Citation:
“Brain-computer interface training fosters perceptual skills to detect errors.” Deland H. Liu, Fumiaki Iwane, Minsu Zhang, Leonardo G. Cohen, and José del R. Millán. Advanced Science; Published Online: July 13, 2026 (DOI: 10.1002/advs.76153).
URL Upon Publication: http://doi.wiley.com/10.1002/advs.76153
Author Contact: Nat Levy, Editorial Manager at the University of Texas at Austin Cockrell School of Engineering, at 512-471-2129 or nat.levy@utexas.edu
About the Journal
Advanced Science is a premier interdisciplinary open access journal covering fundamental and applied research across a broad range of fields, including materials science and chemistry, physics and engineering, life and health sciences, earth and environmental sciences, as well as social sciences and humanities. Advanced Science publishes cutting-edge research through rigorous, efficient, and fair review process, ensuring fast publication with high quality standards and an exceptional author experience. Advanced Science is the flagship journal of Wiley’s Advanced Portfolio: a family of globally respected, high-impact journals that disseminate the best science from well-established and emerging researchers so they can fulfill their mission and maximize the reach of their scientific discoveries.
About Wiley
Wiley is a global leader in authoritative content and research intelligence for the advancement of scientific discovery, innovation, and learning. With more than 200 years at the center of the scholarly ecosystem, Wiley combines trusted publishing heritage with AI-powered platforms to transform how knowledge is discovered, accessed, and applied. From individual researchers and students to Fortune 500 R&D teams, Wiley enables the transformation of scientific breakthroughs into real-world impact. From knowledge to impact—Wiley is redefining what's possible in science and learning. Visit us at Wiley.com and Investors.Wiley.com. Follow us on Facebook, X, LinkedIn and Instagram.
Journal
Advanced Science
Article Title
Brain-computer interface training fosters perceptual skills to detect errors
Article Publication Date
15-Jul-2026
