Classificação de vocalizações em indivíduos com TEA: avaliação de modelos de machine learning

Abstract

This study investigates the application of machine learning techniques for classifying nonverbal vocalizations of minimally verbal individuals with Autism Spectrum Disorder (ASD). Based on the American ReCANVo dataset, six categories of vocalizations were explored: delight, dysregulated, frustrated, request, selftalk, and social. The process involved acoustic feature extraction, class balancing, and training multiple models. The experimental evaluation demonstrated that models trained exclusively on the ReCANVo dataset are ineffective at generalizing to the Portuguese-speaking context. In a test with 53 vocalizations from a Brazilian individual, the 75.47% accuracy achieved by the best model (SVC) proved to be misleading, as it was concentrated on a single over represented class while failing on the minority classes. These results demonstrate the lack of effective cross-cultural generalization and highlight the critical need for data aligned with the local linguistic context. To address this gap, the mobile application VocalizeAI was developed to enable the creation of a Brazilian dataset, which is essential for advancing assistive communication technologies for individuals with ASD.