On Feature Extraction for Voice Pathology Detection from Speech Signals

Reliable, automatic and objective detector of pathological voice disorders from speech signals is a long sought-for tool, by voice clinicians as well as by general practitioners. Such a detector can also be used for low-cost and noninvasive mass-screening, diagnosis and early detection of voice pathology for professionals using voice as an essential career tool, for humans working in risky environments such as chemical factories, and for the general population. Following years of research and significant advancements in voice pathology detection and classification, correct detection and classification rates of various pathology stages are still insufficient for reliable and trusted large-scale screening. The research work in this field generally splits in two stages: first, extraction of meaningful feature sets, and second, using these features for classification of speech recordings into healthy condition and different pathological cases. This work examines the performance of state-of-theart methods, and investigates their weaknesses. Methods examined include features in time, frequency, perturbations, noise and spectral structure; also examined are features more related to the glottal source. Those features are evaluated by different machine learning techniques. This paper describes ongoing work to improve feature sets for more accurate detection of pathology in voice signals, aiming at overcoming certain weaknesses in current state-of-the-art methods, with emphasis on early-stage cases. Promising results on real-world pathologic and healthy samples, recorded in voice clinics, are shown, and future directions discussed.