Music similarity search

Abstract

Technological advances in audio compression and computer networking in the last decade have resulted in an increase of availability of audio documents. Computing music similarity for huge collections is getting more important as is evident in commercial applications. The demand for computational music similarity exists particularly because it enables automatic searching, classification, playlist generation, audio fingerprinting and other tasks that would involve manual or human involvement. This project focuses on extracting features from a music file, namely MP3, storing them, and evaluating the similarity between music files. In this way, the pure content of audio will be analysed as opposed to textual meta-data that might exist. This content based music information retrieval will enable users to search for music based on query by example. Unfortunately, there are no standardised methods for music information retrieval and for computing music similarity that prove to give the correct result with reasonable performance. This is mainly due to the fact that the methods are processing intensive and memory intensive as music files are rather large. Furthermore, music is an art and thus the nature of what is similar is very subjective. The project aims at creating and evaluating low level global acoustic features that would define similarity, to finally produce content based music information retrieval.