Learning to Listen: Matching Song Covers to Original Songs via Supervised Learning Methods

Cover song identification is a trivial problem for human beings. A human can easily identify whether a song is a cover of another or not, without using much effort. However, cover song identification is not an easy task for machines, as the numerical audio data differs significantly. Since a major goal of Machine Learning and Artificial Intelligence is to make machines mimic human-like cognition capabilities, cover song recognition seems to be an interesting problem in the field. This project deals with matching cover songs to the original ones by using features extracted from the Million Songs Dataset, which provides features for about one million songs. Additionally, SecondHandSongs data provides clusters of the original and cover songs grouped together. In this project, we combine these two data-sets to train traditional and slightly modified versions of traditional machine learning algorithms to match cover songs to their original versions.