Exploring the Music Genome: Lyric Clustering with Heterogeneous Features

This research explores the clustering of songs using lyrics features grouped into similar classes and heterogeneous combinations. Simple techniques are used to extract 140 features for analysis with Kohonen self-organising maps. These maps are evaluated using visual analysis and objective measures of validity with respect to the clustering of eight hand-selected song pairs. According to gold standard human-authored playlists, judgments of song similarity are based strongly on music, however this observation may be limited to playlists and is not necessarily extensible to music in the wider domain. In particular, since test song pairs could only be effectively matched when they were from the same genre, analysis of the correspondence between lyrics and expert human judgments of genre and style may be more fruitful than comparison with similarities observed in playlists. Results suggest that for music in the hard-to-differentiate categories of pop, rock and related genres, a combination of features relating to language, grammar, sentiment and repetition improve on the clustering performance of Information Space with a more accurate analysis of song similarity and increased sensitivity to the nuances of song style. SOM analysis further suggests that a few well-chosen attributes may be as good as, if not better than, deep analysis using many features. Results using stress patterns are inconclusive. Although results are preliminary and need to be validated with further research on a larger data set, to the knowledge of this author this is the first time success has been reported in differentiating songs in the rock/pop genre.