Music Boundary Detection Using Local Contextual Information Based on Implication-Realization Model

In this study, we propose a novel melodic boundary detection method using the analysis results of the Implication-Realization (I-R) Model as features of machine learning. Melodic boundary detection is a task for identifying perceptual boundaries inside a note sequence, such as the phrase endings. An input feature that can express melodic expectation is important for detecting perceptual boundaries. Thus, we propose a melodic boundary detection method that incorporates features based on the I-R model, a model of local melodic expectation. To investigate the usefulness of the I-R features, we studied the impact of the I-R features on melodic boundary detection performance. The results showed that the addition of the I-R model improves the boundary detection F-measure by three points, exceeding the previous state-of-the-art.