Análise e classificação das segmentações de músicas

Abstract

The structural analysis of music contains two subproblems: boundary detection and structural grouping, which are analyzed in this work in detail. The result of the first subproblem directly affects the second one. To automate these activities, the Music Structure Analysis Framework (MSAF) was used with two algorithms, Convex Non-Negative Matrix Factorization and 2D Fourier Magnitude Coefficients, which execute these two activities with the k-means clustering technique. In this work, beyond the analysis of MSAF's performance, the music application Moises was used. In this work, a small, but rich in details, dataset was built, which relied on the assessment of a professional musician and the author, a student of theoretical and practical music specializing in electric guitar. To analyze the obtained results, the main metrics were Pairwise Frame Clustering (PWF) and F-measure. A comparative analysis was made between the results of the boundary detections and the suggested labels (structural grouping) of the resulting segmentations of the MSAF framework and the Moises app, which diverge from the golden pattern. The mean of the result of PWF in 5 songs of the dataset for Moises was 42,29%, with MSAF, algorithm 2DFMC, 50,08%. Lastly, for the CNMF algorithm in MSAF, PWF reached 44,18%. It was concluded that the divergences of the notations adopted influenced directly in the final result achieved by the applications, regardless of being the framework or the Moises app, which reinforces the complexity of the activity to handle its subjectivity, besides evidencing the disagreement between the evaluating musicians of the datasets.