Music segmentation analysis in Iberian folk music

Abstract

Music segmentation is one of the fundamental tasks within music information retrieval (MIR). It aims to understand and mimic how our brain understands music and segments it into different structural elements such as beats, meters, phrases, and sections. This work focuses on the automatic phrase segmentation subtask. The models used for phrase segmentation can be divided into three major categories: rule-based, supervised, and unsupervised models. Rule-based models are based on musical rules and how humans perceive music phrases. The second and third categories handle phrase detection as a machine-learning problem and detect the musical phrases by analyzing the pieces as an annotated or unannotated data. Among many statistical and music-theory approaches, the local boundary detection model (LBDM) is a prominent method due to its accuracy and cognitive-basis explainability. It relies on musical changes to shape a novelty curve and find the candidate segments based on pitch intervals, inter-onset intervals, and rest information extracted from symbolic musical surface structure (i.e., scores). An identified limitation of the LBDM algorithm is the lack of repetition awareness in the musical surface structure. To this end, we study the enhancement of the baseline model with structural repetition information resulting from the variational Markov oracle (VMO). A finite-state automaton that can efficiently search for factors (substrings) in a string. Furthermore, we assess optimal weights for each of the surface elements information extracted, using genetic algorithm search heuristic and entropy, which is a measure of information in a desired variable change. To evaluate our models, we adopted a novel dataset of Iberian folk music consisting of 59 Portuguese and 738 Spanish folk pieces in symbolic representation, and we used the commonly used F-measure as our reference metric to assess the effectiveness of our models. Our results show that with a good amount of good-quality data and an optimization algorithm, the LDBM, or any weight-based model, can be optimized and achieve better performance for a specific genre, region, or dataset. Our model, with sufficient training data, managed to increase the overall F-measure of the original LBDM in Spanish folk music from 60% to 67% which is even higher than what has been reported in previous literature.