Voice pathologies identification speech signals, features and classifiers evaluation

Abstract

Voice pathology identification using speech processing methods can be used as a preliminary diagnosis. This study implements a set of identification systems to screen voice pathologies using voice signal features from the sustained vowel /a/ and continuous speech. The two signals tasks are evaluated using three acoustic features applied to four classifiers. Three main classes are identified: physiological disorders; neuromuscular disorders; and healthy subjects. The main objective of this work is to evaluate which voice signal is more reliable for voice pathology diagnosis, which acoustic feature has more pathology information and which is the best classifier to carry out this task. The best overall system accuracy is 77.9%, obtained with Mel-Line Spectrum Frequencies (MLSF) feature extracted from continuous speech and applied to a Gaussian Mixture Models (GMM) classifier.