The prospect of decoding speech by tapping into neural activity has become a reality thanks to advanced machine learning models. But recently, it’s been taken to the next level due to technology developed by researchers from HSE University and the Moscow State University of Medicine and Dentistry. Indeed, they’ve devised a model that can effectively predict the next word spoken by an individual by connecting it to a small set of minimally invasive electrodes that have been implanted in a limited cortical area to track neural activity. The findings of this project, financed through a grant from the Russian government’s ‘Science and Universities’ National Project, were published in the Journal of Neural Engineering.
While there is abundant technology for improving communication functions in patients, such as “silent speech” platforms that interpret speech intention by tracking articulatory muscle movement, they leave out a fair portion of patients with more debilitating disorders like facial muscle paralysis. The incorporation of a brain-computer interface (BCI) in speech neuroprostheses does much to widen the range of patients whose lives can be improved through the technology.
The current study from the HSE Centre for Bioelectric Interfaces and the MSUMD worked with data from two epileptic patients who already had intracranial electrodes implanted for presurgical mapping purposes. They were tasked with reading six 26-word sentences aloud, with the phrasing varying in structure and most of the words within the sentence beginning with the same letter. The electrodes gauged the subject’s brain activity during these readings, then aligned the results with audio signals in 27 classes. This training dataset was fed into a model with a neural network-based architecture that managed to achieve 55% and 70% accuracy in the two subjects. That level of predictive accuracy puts the tech in good company with that of similar implanted-electrode studies, but in comparison those required the electrodes to cover the entire cortical surface.