Building ensembles of audio and lyrics features to improve musical genre classification

Digital audio has become an almost ubiquitously spread medium, and for many consumers, digital audio is the major distribution and storage form of music. Numerous on-line music stores account for a growing share of record sales. The widespread adoption of digital audio on home computers and especially mobile devices, and numerous on-line music stores show the size of this market. Handling the ever growing size of both private and commercial collections however becomes increasingly difficult. Computer algorithms that can understand and interpret characteristics of music, and organise and recommend them for and to their users can be of great assistance. Music is an inherently multi-modal type of data, and the lyrics associated with the music are as essential to the reception and the message of a song as is the audio. Album covers are carefully designed by artists to convey a message consistent with the music and image of a band. Music videos, fan sites and other sources of information add to that in a usually coherent manner. In this paper, we focus on exploring the lyrics domain of music, and how this information can be combined with the acoustic domain. We evaluate our approach by means of a common task in music information retrieval, musical genre classification. Advancing over previous work that showed improvements with simple feature fusion, were we successfully demonstrated simple approaches of combining different representations of music, we apply a more sophisticated machine learning technique, ensemble classification. The results show that the approach is superior to the best choice of a single algorithm on a single feature set. Moreover, it also releases the user from making this choice explicitly.