Literary review of content-based music recognition paradigms

Abstract

During the last few decades, a need for novel retrieval strategies for large audio databases emerged as millions of digital audio documents became accessible for everyone through the Internet. It became essential that the users could search for songs that they had no prior information about using only the content of the audio as a query. In practice this means that when a user hears an unknown song coming out of the radio and wants to get more information about it, he or she can simply record a sample of the song with a mobile device and send it to a music recognition application as a query. Query results would then be presented on the screen with all the necessary meta data, such as the song name and artist. The retrieval systems are expected to perform quickly and accurately against large databases that may contain millions of songs, which poses lots of challenges for the researchers.

This thesis is a literature review which will go through some audio retrieval paradigms that allow querying for songs using only their audio content, such as audio fingerprinting. It will also address the typical problems and challenges of audio retrieval and compare how each of these proposed paradigms performs in these challenging scenarios.