Automatic tune search

Abstract

Conventional music retrieval systems focus on retrieving musical scores based on attributes such as the title and composer. These labels, however, by no means describe the content of a song and a tune can exist independently of them. It is commonplace to come across a person who remembers part of a tune and can easily hum it and yet is unable to remember the name of the tune. This dissertation explores the idea of retrieving musical scores based solely on their content. It reviews the methods which have been used in previous systems and proposes a variation of one of these methods to make use of rhythmic information. This is based on the assumption that some people may be more capable of reproducing a rhythm rather than perfect pitch. This theory is borne out to some degree by the results of this work since certain subjects participating in the experiments were able to reproduce reasonably correct rhythm whilst they were completely out of pitch. Our system, called the Automatic Tune Search, extracts both pitch and rhythmic information from standard MIDI files to populate a musical database. It also extracts such information from hummed queries input by the user, taking into consideration the errors which may be generated from imperfect recollection of a tune and limited singing ability of the user. An approximate string matching algorithm was implemented to gauge the similarity between two songs and graded results are displayed. The experiments carried out have given encouraging results and point toward the importance of the rhythm as a distinguishing musical feature.