Classificação de Distúrbios Vocais Utilizando Aprendizado de Máquina e Análise Espectral

Abstract

This work proposes a machine learning-based approach for detecting voice disor ders, utilizing acoustic features extracted from audio, such as Mel-Frequency Cepstral Coefficients (MFCCs) and voice perturbation parameters, such as jitter and shimmer. Additionally, a comparison was made using data augmentation to expand the dataset and improve the model’s accuracy. The adopted methodology divides the signal into 16 frequency bands for feature extraction. After extraction, each frequency band is used to train a machine learning model, totaling 16 models. The results from each individual model are then combined into a meta-model, responsible for determining the final classifi cation of the audio. In the data augmentation scenario, the trained meta-model achieved an accuracy of 82.33%, while one of the individual models reached an accuracy of 89.16%. The approach proved to be effective, as it surpassed convolutional neural network-based methods, even when using fewer resources.