Identificação de gênero musical por meio de redes neurais convolucionais

Abstract

Musical classification is a collection of problems within the field of Musical Information Retrieval (MIR), including: identification of instruments present in the music, identification of the set of notes, identification of musical genre, decomposition of sound sources, and others. In this thesis, the problem of musical genre classification will be addressed, and the implementation of a Convolutional Recurrent Neural Network (CRNN) will be studied, with its performance compared to Convolutional Neural Networks (CNN). Additionally, the architecture of the networks will be built with musical information in mind, i.e., we will be utilizing characteristics present in musical signals to try to improve the performance of neural networks in the classification problem. The convolutional network structure with recurrent layers, together with enhancements in the design of the architecture of the networks considering the structure of the signal, is able to achieve a categorical accuracy of 77% and an AUC-ROC of 0.96 in the problem of identifying musical genres, comparable to the state-of-the-art.