Dificuldade no julgamento perceptivo-auditivo da qualidade vocal : uma abordagem baseada em aprendizado de máquina

Abstract

Auditory-Perceptual Judgment (APJ) of voice is widely used in clinical practice and speech-language pathology training, but it presents high variability among judges, which compromises its reliability. Traditionally, studies focus on factors related to the judge or the speech task, neglecting the possibility that the voice itself presents different degrees of difficulty for judgment. The objective of this work is to develop a pattern classification system to predict the difficulty of the APJ of voice based on ML models trained with concordance data and acoustic measurements. The voices analyzed in this study belong to the database of the Integrated Laboratory of Voice Studies (LIEV) at the home institution and are frequently used in research developed by the group. Initially, clustering techniques were applied to group 295 voices according to the degree of APJ difficulty, considering the APJ coefficient of variation and fuzzy concordance among five expert voice judges. The K-means algorithm was selected because it presented the best internal validity index (Calinski-Harabasz) and the best visual separation between groups. The voices were classified into three difficulty levels: easy, medium, and difficult. Subsequently, fourteen supervised classification learning models were trained to classify, based on acoustic measurements, the degree of voice judgment difficulty. The Multinomial Regression, Multilayer Perceptron, and Random Forest models performed best for different vocal parameters. The findings demonstrate the potential of combining unsupervised and supervised methods to classify JPA difficulty, contributing to the creation of more objective and evidence-based training systems. The proposal provides support for the development of training simulators with progression based on the perceptual difficulty of voices, with a direct impact on training and clinical practice in Speech-Language Pathology.