Reconhecendo gêneros musicais brasileiros similares por abordagem baseada em segmentos e aprendizado profundo usando transformadores visuais

Abstract

This research investigates the use of self-attention mechanisms in visual transformer architectures, focusing on the Vision Transformer (ViT) model for musical genre recognition and similarity analysis. The study involves the construction of the BYRM Dataset, a curated collection of Brazilian regional music comprising 1.082 tracks across ten culturally diverse genres, automatically extracted from YouTube album videos. The dataset supports controlled experiments using multiple excerpts per track, varying both the temporal position and the segment duration. The ViT model was trained using Melspectrograms as input and evaluated with metrics such as accuracy, precision, recall, and F1-score. The best configuration achieved 81,94% accuracy and 81,84% F1-score. Beyond classification, this dissertation also proposes an analysis of the vector representations learned by the model, applying dimensionality reduction techniques (PCA, t-SNE,and UMAP) and quantitative similarity measures, such as cosine distance, to investigate the proximity between genres in the latent space. Results indicate that ViT successfully captures meaningful stylistic relationships, grouping similar genres such as samba and pagode or vaneira and xote gaúcho. This work contributes to the field of Music Information Retrieval (MIR), demonstrating the potential of attention-based models in the classification and interpretation of musical genres in culturally rich and diverse contexts.