Music Retrieval based on Melodic Similarity

Abstract

This thesis introduces a method for measuring melodic similarity for notated music such as MIDI files. This music search algorithm views music as sets of notes that are represented as weighted points in the two-dimensional space of time and pitch. Two point sets can be compared by calculating how much effort it would take to convert one into the other; effort is measured by determining how much weight has to be moved over what distances. To make these point set comparisons efficient enough for searching large databases, the distances between every item (point set) in the database and a small, ?xed set of special (vantage) point sets can be pre-calculated. Whenever a new query needs to be compared to the items in the database, one can restrict the search to those items with similar distances to the special point sets. For studying the performance of the transportation-based search algorithm and other, similar ones, the creation of a ground truth for a large music collection (RISM) is described, along with a performance measure and the application of both the ground truth and the measure for the MIREX algorithm competition.