A Novel Local Alignment-Based Approach to Motif Extraction in Polyphonic Music

The paper provides a novel approach to musicologically-informed intra-opus motif detection within polyphonic music scores. We extract diatonic interval sequences from each voice of a score; sequence segmentation is performed via pairwise local alignment between each pair of voices. From the output of this step, string-based approaches are used for motif discovery. Specifically, a weighted directed acyclic graph is constructed, giving a custom measurement of motif importance. A selection and filtration procedure is applied according to a set of rules and music structural information, to generate a final selection of music motifs. The ground truth annotated JKUPDD dataset is used for evaluation of the proposed methodology. The results demonstrate that this algorithm is capable of extracting musically meaningful motifs with high precision and recall.