In order to detect laryngeal pathologies by acoustic analysis of the speech wave, several acoustic features have been proposed in literature. The evaluation of their individual performance in discriminating between normal and pathological speakers has been done either qualitatively by inspection or by classical statistical analysis. It seems that the evaluation of a whole set of acoustic features has never been done without explicit reference to a decision model. On the other hand there are strong indications that the distributions underlying the features space are multimodal. In order to evaluate quantitatively their discrimination power without reference to a statistical model, we performed a clustering analysis on a set of six well-known acoustic features. The features were computed from the acoustic signal of the steady vowel /a/ uttered by 37 normal speakers and by 24 dysphonic speakers. The results confirm the good performance of the pitch and amplitude perturbation measures reported elseewhere. On the contrary, the acoustic features solely defined on the residue signal show poor discrimination ability. © 1982.
Schoentgen, J. (1982). Quantitative evaluation of the discrimination performance of acoustic features in detecting laryngeal pathology. Speech Communication, 1(3–4), 269–282. https://doi.org/10.1016/0167-6393(82)90020-6