Bridging the Semantic Gap in Music

Musical genres are categorical labels intended to group similar sounding musical tracks together. Precise labeling of genre has the advantage of better connecting listeners to music they will enjoy. To this end, we have gathered both low-level features derived from MP3 files, and high-level features assigned by humans—such as genre or how suited a song is for dancing—which we intend to predict. Through machine learning, we aim to discover patterns in the low-level features which then can be used to make predictions of the high-level features. Lidy and Rauber (2005) used the low-level features of Statistical Spectrum Descriptions and Rhythm Histograms to predict a song’s genre with 75% accuracy. Those features are included in our dataset. Our findings explore the nature of high-level classification by answering the question: which combinations of low-level features, paired with which machine learning techniques, are able to most accurately predict high-level musical features.