Kya Masoumi – August 18th, 1:30 pm, Rowe 4025
Title: Enhancing Personal Communication: A Brain-Computer Interface Integrated With Machine Learning For Intent-Based Speech Decoding
Abstract: Brain-computer interfaces offer a possible communication channel for people whose speech is compromised by conditions such as amyotrophic lateral sclerosis, stroke, or cerebral palsy, on the premise that the neural planning of words survives the loss of the motor pathway that would articulate them. This thesis asks whether that planning activity is recoverable from non-invasive electroencephalography. Ten neurologically unimpaired adults produced five target words, cued as written text or as images, either spoken aloud or silently imagined, and again with each target embedded in a carrier sentence. Epochs were locked to the moment of production and classified by four approaches drawn from distinct modelling families: a support vector machine, shrinkage-regularised linear discriminant analysis, a Riemannian geometry classifier, and a convolutional neural network.
Word identity was decodable above the 20% chance level whenever the word was spoken aloud, reaching 38.3% and holding for every classifier in both cue modalities. Imagined production was not reliably decodable, with seven of eight conditions failing to exceed chance. Classifiers trained on isolated words did not detect the same words inside sentences, one of thirty-two transfer pairings surviving correction. Cue modality had little effect. The convolutional network was outperformed throughout. Three design confounds qualifying these findings are documented.