One Channel Subvocal Speech Phrases Recognition Using Cumulative Residual Entropy and Support Vector Machines

This paper presents the design and implementation of a subvocal speech pattern recognition system using only one EMG channel. The objective of the system is, after being trained, to identify and classify limited-vocabulary sets of speaker-dependent Spanish words. First we present the EMG signal acquisition board designed and constructed for this end. Then, we describe the preprocessing stage where denoising and activity detection occurs. Then the various feature spaces representations alongside the different candidate classifiers are explained and compared; we obtained the best results using a filter bank analysis followed by cumulative residual entropy (CRE) profile and a Support Vector Machine (SVM) classifier. For testing, we considered two possible application of this type of systems: confidential communications and voice recognition in high acoustic noise environments. For both a vocabulary made up of six words was tested, and the latter was tested while simulating fire noise and also compared to a vocal speech system. The performance of both applications was evaluated on two groups of four-subject with no speech disorders, obtaining mean F₁-Scores of 91.32 % and 90.83 % respectively.