SONBRA: uma base de dados para a classificação automática de gêneros musicais brasileiros

Abstract

After reviewing the most relevant musical databases available for studying music and musical genres in the field of Music Information Retrieval, a scarcity of databases containing Brazilian musical genres was observed. Therefore, this study aimed to create the SONBRA, a stratified database with the following musical genres: Bossa Nova, Forró, Forró Piseiro, Funk, Pagode, Samba, Samba-Enredo and Sertanejo. In order to increase the number of samples for each genre, each track was divided into five fragments, and the following features were extracted from each fragment: Fourier Tempogram, Tempogram, Mel, MFCC, Chroma STFT, Chroma CQT, Chroma CENS, Tonnetz, ZCR, Spectral Centroid, Spectral Roll Off, Spectral Bandwidth and RMS. To assess whether the database is suitable for genre classification, it was evaluated using classification models: KNN, MLP, XGBoost, Decision Tree, Random Forest, SVM, and an ensemble approach, specifically a Voting system. The best result was achieved with the Voting model, which obtained 83.2% accuracy. Among the individual models, XGBoost performed the best with 82.5% accuracy, while the Decision Tree model had the lowest performance, achieving 62.7% accuracy. Overall, the genres that the models classified most easily were Forró Piseiro, Funk, and Samba-Enredo, with Funk standing out. In contrast, the genre that exhibited the greatest difficulty in classification was Sertanejo. Regarding the extracted features, the models performed best with combinations of Mel, MFCC, Tempogram, and Chroma CENS. Therefore, it was demonstrated that the database and the extracted features can be used for genre classification tasks.