Duke Researchers Develop More Scalable Brain-to-Speech Technology

Stroke, ALS, and other neurodegenerative and acute brain disorders can rob people of their ability to speak. Although brain-to-computer interfaces can restore this capacity by translating brain activity into machine-generated speech, their widespread and practical availability has been limited by the need for patient-by-patient data collection. To overcome this major barrier, Duke researchers are making these systems faster and easier to deploy.

In a study published in Nature Communications, the research team demonstrated for the first time in humans that pooling data across multiple patients can work, rather than relying entirely on fully individualized training. Led by Greg Cogan, PhD, associate professor in neurology, and Jonathan Viventi, PhD, Hawkes Family Associate Professor in Biomedical Engineering, the team utilized high-resolution recordings (micro-ECoG recordings) from awake neurosurgical patients. By aligning brain activity across individuals, they successfully trained a shared decoder, which improved performance while decreasing the amount of data required from each patient.

According to Cogan, a functional decoder in their study could be built using as little as five minutes of a new patient's own recordings, supplemented by aligned data from other individuals. This advancement represents an encouraging step toward making these brain-to-speech systems more scalable and accessible. The study's co-authors from Duke include Zac Spalding, Suseendrakumar Duraivel, Shervin Rahimpour, Charles Wang, Katrina Barth, Ceci Schmitz, Shivanand P. Lad, Allan H. Friedman, and Derek G. Southwell. Funding for the research was provided by the National Institutes of Health, the National Science Foundation, and a Duke Institute for Brain Sciences Incubator Award.