EEG-Based Recognition of Silent and Imagined Vowels

Abstract

This work proposes a framework for future Silent Speech Interfaces (SSI) based on non-invasive EEG recordings. Specifically, the information embedded in the brain signals related to the production – overt, covert and imagined production – of the Italian vowels /a/ and /i/ allowed to distinguish the vowels relying on discriminative features calculated by the Ambiguity Function in the context of time-frequency analysis, and ranked by the Fisher contrast. The vowels were classified by using a multilayer feed-forward ANN. Overall, intra-subject classification accuracies, as measured by the area under the ROC curve, ranged from 0.84 to 0.96 for overt production, from 0.83 to 0.96 for covert production, and from 0.89 to 0.98 for imagined vowels. Results indicate significant potential for the use of speech prosthesis controllers for clinical and military applications.