UTHealth Houston researchers develop technology that can decode speech across different patients

UTHealth Houston researchers develop technology that can decode speech across different patients

Written by: Catherine Marfin | Updated: October 10, 2025

A new study led by researchers at UTHealth Houston has brought scientists one step closer to creating a more efficient device that can understand and translate a person's speech directly from their brain.

The results were published in Nature Communications.

Prior to the team's research, the technology - known as a brain-computer interface - was limited in that it required patients to spend hours or weeks training the technology and needed recordings from specific, intact brain regions to work accurately. That's not always feasible for patients who have already lost the ability to speak, a condition known as aphasia, due to stroke, brain injury, or other disease.

Brain-computer interfaces work by reading a person's brain signals during attempted speech and turning them into words, either as text on a screen or through a voice synthesizer.

To make the technology more efficient, principal investigator Nitin Tandon, MD, professor of neurosurgery at McGovern Medical School at UTHealth Houston and director of the Texas Institute for Restorative Neurotechnologies at UTHealth Houston, and his team used cross-subject transfer learning, a technique where a model trained on one person's brain data adapts to data from another person, instead of starting over from scratch with each patient.

Researchers recorded brain activity from 25 epilepsy patients with depth electrodes, or thin, surgically implanted devices that monitor brain waves, while patients spoke challenging tongue twisters. The brain-computer interface was then able to translate the brain activity into phonemes, the smallest units of sound, like "p" or "sh."

"With it being a complex tongue twister task, your speech system is at a heightened alertness to minimize errors," said Tandon, who is director of the Texas Comprehensive Epilepsy Program, as well as the Nancy, Clive and Pierce Runnells Distinguished Chair in Neuroscience of the Vivian L. Smith Center for Neurologic Research and BCMS Distinguished Professor in Neurological Disorders and Neurosurgery at McGovern Medical School. "This means there is maximum engagement of the speech system and a lot of neural activity that we can tap into to decode what they are saying or trying to say."

Researchers used the data from the sessions to build a shared "language" of brain signals that allowed the brain-computer interface to be fine-tuned with a small amount of data from each new individual.

Even when a person had limited brain coverage or only a short recording session, the shared model still decoded speech more accurately than training a model from that person alone, according to the study.

While brain-computer interface technology has rapidly advanced over the last several years, the research shows that by leveraging data from many people, future brain-computer interfaces could work reliably even for new patients with minimal amounts of data or damaged speech brain regions.

"This allows us to create this library that you can read from when you have somebody with a brain injury to try to replicate normal language," Tandon said. "?This is a really foundational step in being able to help people with aphasia."

Aditya Singh, MD, PhD candidate in The University of Texas MD Anderson Cancer Center UTHealth Houston Graduate School of Biomedical Sciences, served as lead author of the study.

Additional co-authors on the study are Tessy Thomas, PhD, of Johnson & Johnson MedTech; Jinlong Li, PhD, of Rice University; Greg Hickok, PhD, of University of California, Irvine; and Xaq Pitkow, PhD, of Baylor College of Medicine.