Evaluating the Genre Classification Performance of Lyrical Features Relative to Audio, Symbolic and Cultural Features

This paper describes experimental research investigating the genre classification utility of combining features extracted from lyrical, audio, symbolic and cultural sources of musical information. It was found that cultural features consisting of information extracted from both web searches and mined listener tags were particularly effective, with the result that classification accuracies were achieved that compare favorably with the current state of the art of musical genre classification. It was also found that features extracted from lyrics were less effective than the other feature types. Finally, it was found that, with some exceptions, combining feature types does improve classification performance. The new lyricFetcher and jLyrics software are also presented as tools that can be used as a framework for developing more effective classification methodologies based on lyrics in the future.