Envelhecimento de características da voz

Abstract

This work presents an approach based on linear models to recover the vocal identity of individuals with dysphonia, using recordings made before the onset of the disorder. The method applies transformations aimed at approximating the representation of vocal age to what the individual would have had if the disorder had not compromised voice quality. The classification models achieved a mean absolute error of 3.49 years in age prediction, although the analysis was limited by an imbalanced dataset with low representativeness across different age groups. Furthermore, a simple linear model applied to speaker embeddings showed promising performance within the scope of the available data. Despite the demonstrated potential, further validation with larger and more diverse datasets is needed to ensure its practical applicability and adaptation to individual particularities.