SYLLABLE-BASED SPEECH RECOGNITION USING ELECTROMYOGRAPHY AND DECISION SET CLASSIFIER

Abstract

During the speech, contractions of muscles in the speech apparatus produce myoelectric signals that can be picked up by electrodes, filtered and analyzed. The problem of extraction of speech information from these signals is significant for patients with damaged speech apparatus, such as laryngectomy patients, who could use speech recognition based on myoelectric signal classification to communicate by means of the synthetic speech. In the most previously conducted research, classification is performed on a ten word vocabulary which resulted in a good classification rate. In this paper, a possibility for myoelectric syllable based speech classification is analyzed on a significantly larger vocabulary with novel decision set based classifier which is simple, easy to adapt, convenient for research and similar to the way humans think. In order to have a high quality of recorded myoelectric signals, analysis of the optimal position of electrodes is performed. Classification is performed by comparison between syllable combination and whole words. Based on classification rate, words can belong to easy, medium or hard to distinguish group. Results based on generated list of best matching combinations show that decision set analysis of myoelectric signals for speech recognition is a promising novel method.