A Review on Automatic Assessment and Detection of Pathological Speech

Communication through speech is an essential requirement for social participation. It is a way through which people exchange information, emotions, ideas and feelings. Hence making it the most popular form of communication. Speech processing (SP) research has primarily focused on modelling the activity of voice sources and the activities in the oral cavity. However, SP inherently corresponds to other physiological activities, such like respiration and heart activity, which changes due to several reasons like neurological speech disorder or motor speech disorder (e.g. Dysarthria), mood/emotion, neurodegenerative diseases (e.g. Parkinson’s disease), social setting, environmental setting (e.g. loud/quiet environment). These changes, in turn, can influence speech communication. Changes in pathologies impair the speech construction process, thus producing decreased speech intelligibility and communicative ability. Automatic pathological speech (APS) assessments are indispensable in assisting the clinical diagnosis, treatment and management of speech disorders. In contrast to clinicians’ subjective and time-consuming auditory-perceptual analyses, such automatic assessments provide accurate, objective and practically usable assessments. APS distinguishes between normal and pathological speech and speech intelligibility assessment (SIA) estimating intelligibility as the ratio of the number of words correctly comprehended by human listeners to the total number of words. In this paper, discussion about the techniques used in extracting features of pathological speech (PS) along with the classical machine learning approaches or deep learning approaches used for APS detection. Comparative analysis on various datasets is done for comprehending the anatomy of PS. Brief discussion on APS and SIA will be given. As these are two significant automatic analyses for a possible future tool that can be used for medical purposes.