Modified graph spectra approach for music information retrieval

Abstract

For large collections of melodies, it is too slow to calculate the distance of the query object from all the test objects. A reasonably fast method is needed to coarse sieve and identify a set of closest tunes .One way to achieve this is by using nearest neighbour methods on indexed database. Graph Spectra [1] Indexing method, which index tunes using their Laplacian spectra have proved successful in finding nearest neighbours of a query melody. In this paper we propose a modified version of Graph Spectra indexing method called Ratio Graph Spectra. A comparison of commonly used geometric methods such as Earth Movers Distance (EMD) and Indexing methods specifically Graph Spectra with Ratio Graph Spectra is conducted. We observed that R Graph Spectra indexing method outperformed other methods giving up to 80 % accuracy in retrieving nearest neighbour for a query tune. © 2011 Springer-Verlag Berlin Heidelberg.