Phrase Segmentation of Jiangnan Music Using YNote Representation

This study aims to develop an interpretable and effective model for automatic musical phrase segmentation of Jiangnan music. The research utilizes a corpus of 183 Jiangnan musical pieces, manually annotated by experts and represented in YNote, a notation system designed for machine learning that transforms music into structured text. The core methodology involves a statistical probability model that calculates the likelihood of a phrase boundary based on the combinations of “duration” and “pitch” of the notes immediately preceding and following a potential break. To validate the model, probability rules were derived from a training set of 30 pieces and tested on the remaining 153 pieces. Experimental results demonstrate that duration is the most critical feature for identifying phrase boundaries. A model relying solely on duration information achieved an F1-score of 0.6924 under a uniform sampling strategy. These findings confirm that local note features are sufficient for effectively identifying phrase structures in the rhythmically regular context of Jiangnan music. This research not only proposes a simple and transparent segmentation method but also validates the potential of YNote in musical structure analysis, providing a foundation for future studies combining interpretable AI with music theory.