USO DE VISUALIZAÇÃO DE INFORMAÇÃO PARA INVESTIGAÇÃO DA CLASSIFICAÇÃO SUPERVISIONADA DE GÊNEROS MUSICAIS

Abstract

This work investigates the application of Artificial Neural Networks (ANNs) for classifying music into different genres using the GTZAN dataset. The primary focus is on leveraging Mel-Frequency Cepstral Coefficients (MFCCs) as features for training and testing a Multilayer Perceptron (MLP) model, combined with information visualization techniques to analyze the training process. The study demonstrates that using 15 MFCC coefficients results in the highest accuracy of 99.25%, effectively distinguishing between musical genres. Additionally, multidi- mensional visualization using the t-SNE technique provided a detailed understanding of how the model processes various audio features, aiding in pattern identification and training improve- ments. The results highlight the ANN’s ability to generalize across different audio inputs and its potential application in music recommendation systems. The findings underscore the relevance of parameter tuning, feature extraction techniques, and visualization methods to improve the performance and interpretability of neural network models in music genre classification.