|
Tree representation and efficient search for melody recognition AbstractThe search of a given melody in large data-bases is one of the problems in the modern topic of music information retrieval (MIR). A huge amount of music files in symbolic formats can be found today in the Internet, and this has motivated new challenges for identification and categorization of music data. A number of pattern recognition techniques can be used to solve this problem. In this paper we explore the capabilities of trees to provide an expressive representation of music information. Trees are compared to string representations in terms of dissimilarity measures, using edit distances. The high computational cost of tree edit distances needs of complexity reduction techniques to be applied. Partial tree edit distances will be considered in order to solve this problem. Also, a new approximate nearest neighbour search for non-vector representation of patterns (such as trees) is applied to speed up the classification. The combination of both techniques produces a significant reduction in classification error rates of string representations while keeping similar classification times.
[Edit] |
