Extracting syllables from EEG signal of speech-imagery

Abstract

Speech imagery recognition from Electroencephalogram (EEG) is one of the challenging technologies for non-invasive brain-computer-interface (BCI). In this report, firstly seventeen syllables appeared in ten Japanese digits are extracted from continuously imagined speech by hand-labelling and evaluated to classify three syllable-groups using Subspace Method (SM). Then, an unlabeled data-set of seventeen short-syllables is collected and tested using 2D-Convolutional NN (CNN). The noise reduction including event related potentials of a prompt pure-tone (ERPs) and the extraction of space-patterns of twenty-one electrodes are also described.