A generalized graph-spectral approach to melodic modeling and retrieval

The development of both musicologically based and efficient music information retrieval metrics to query large music database is crucial in modern music information retrieval, knowledge management and database research. Graph spectral representation of pitch class sequences has proved to outperform other pitch class based melodic similarity methods. Here we compare different spectral approaches to structural queries in databases of symbolic music, which exploits mathematical music theory results to improve the descriptive power of representative graphs. In particular, we explore graph representation of other relevant music features like intervals. The experiments have been conducted on a subset of the RISM collection, and results have been evaluated against a ground truth for the same collection developed for the MIREX competition.