M-tree index for music search based on similarity of cosine contours and tags

Abstract

The similarity between songs is an important part of the MIR (Music Information Retrieval), but the definition of the similarity is very subjective. Similarity can be used, for example, in recommendation systems. These systems recommend similar songs based on the user history from a database. To find similar songs in a fast way we have to index the data. Most database indexes are created for exact item searches. GiST (Generalized Search Tree) gives us the possibility to create an index with a distance function between items. These distances can be used for similarity measures. In this paper, we show how can use music similarity for distance in M-tree, which is a distance-based index. Two similarity metrics are used to create an index of music data: song tags and cosine contour.