Classificação automática de música utilizando aprendizagem de padrões de votação

• Main Author: Shinohara, Vítor Yudi
• Publication Date: 2018
• Format: Bachelor thesis
• Language: por
• Source: Repositório Institucional da UTFPR (da Universidade Tecnológica Federal do Paraná (RIUT))
• Download full: http://repositorio.utfpr.edu.br/jspui/handle/1/6008
• topic: Aprendizado do computador; Música; Redes neurais (Computação); Machine learning; Music; Neural networks (Computer science); CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

Research in MIR have proposed many automatic genre classification systems using machine learning. In this context, two main approaches have been used to describe music tracks: Single Vector Representation (SVR), which uses a single vector, and Multiple Vector Representation (MVR), which uses multiple vectors. For training MVR models, the track is divided into auditory textures, which are all labeled with the ground truth and presented independently during training. When testing, each texture is classified independently, and then some voting scheme must be used to assign a final label to the entire track. When only the votes are available for each texture, the only option available to assign a final label to the track is majority voting. Other techniques which rely on class probabilities for each texture requires those probabilities to be computed by the classifier during prediction. These probabilities are costly to compute, thus are not always available. We present two novel voting schemes as alternatives to majority voting. Both methods use voting pattern learning where only the predicted class of each texture is known, not class probability distributions. The votes were combined in two different ways: a voting histogram and a vote sequence vector. The histograms were used as feature vectors for both K-Vizinhos Mais Próximos (K-NN) and Máquina de Vetores de Suporte (SVM) classifiers. The vote sequence vectors were used as inputs to sequence modelling with Hidden Markov Model (HMM) and two recurrent neural network architectures. The accuracy and accuracy standard deviation of the classification were computed for both proposed methods. The performance of both were compared to the results of the majority voting technique. Student’s T-Test was used to evaluate which gains were statistically significant. The results show that voting pattern learning is relevant and can provide statistically significant performance gains in some data sets.