Music Structure Analysis for Cover Song Identification

Abstract

This master thesis proposes to analyze the structure of audio tracks in order to extract some structural information. The latter information should then be used to characterize the songs in order to retrieve cover songs in a database. The goal of this thesis is to provide some improvement in terms of efficiency and runtime in the cover song identification domain.

This work is split into two parts. First, we analyzed each songs of the cover80 database and given some information retrieved from the audio files, we have characterized each song with his musical pattern.

Second, we used this special characterization of the songs to perform some cover song identification. We present two different solutions to perform this task. One by using the relative position of important parts of the song, such as chorus, and by comparing these positions with the rest of the database, and one by comparing the musical content (in this case the chromas) of these same parts with the rest of the database.

It will be shown that these methods give some interesting results in terms of efficiency and the improvements made in terms of execution time. It allows us to discuss about the trade off that can made between efficiency and runtime.

Finally, we compare our methods with some existing ones and discuss the future of the methods on larger databases and the future of cover song identification.