Electroencephalographic Signals Analysis for Imagined Speech Classification

Abstract

This work aims to interpret the EEG signals associated with actions to imagine the pronunciation of words that belong to a reduced vocabulary without moving the articulatory muscles and without uttering any audible sound (imagined or unspoken speech). Specifically, the vocabulary reflects movements to control the cursor on the computer, and consists of the Spanish language words: "arriba", "abajo", "izquierda", "derecha", and "seleccionar". To do this, we have recorded EEG signals from 27 subjects using a basic protocol to know a priori in what segments of the signal a subject imagines the pronunciation of the indicated word. Subsequently, discrete wavelet transform (DWT) is used to extract features from the segments. These are used to compute relative wavelet energy (RWE) in each of the levels in that EEG signal is decomposed and, it is selected an RWE values subset with the frequencies smaller than 32 Hz. Then, these are concatenated in two different configurations: 14 channels (full) and 4 channels (the channels nearest to the brain areas of Wernicke and Broca). The following three classifiers were trained using both configurations: Naive Bayes (NB), Random Forest (RF) and support vector machines (SVM). The best accuracies were obtained by RF whose averages were 60.11% and 47.93% using both configurations, respectively. Even though, the results are still preliminary, these are above 20%, this means they are more accurate than chance for five classes. Based on them, we can conjecture that the EEG signals could contain information needed for the classification of the imagined pronunciations of the words belonging to a reduced vocabulary.