Silent speech decoder using adaptive collection

Abstract

We investigated a classification method using brain computer interfaces (BCIs) for silent speech. Event-related potentials (ERPs) obtained when four subjects imagined the vocalization of two Japanese vowels while they remained silent and immobilized were recorded. We used an adaptive collection (AC) that adaptively selects suitable output signals of common spatial patterns (CSP) filters and its time duration for classification. The classification accuracies (CAs) were 73-92% for the pairwise classification /a/ vs. /u/ in the use of 63 channels and significantly better than previous study.