Labeling data and developing supervised framework for hindi music mood analysis

摘要

Digitization has created a wide platform for music, in the form of televisions, desktops and other hand held devices. This has increased the reach of musical content as well as its impact on people. Music is often associated with distinct emotional content, generally referred to as music mood. Literature focusing on analyzing the content of a music piece, often discusses music mood as an important metadata. The present article addresses the issue of Hindi music mood classification by considering important issues like taxonomy development, annotation and automated mood classification. We annotated a total of 1540 music clips of 60 seconds duration each, with either of a proposed set of five mood classes derived from Russell’s circumplex model. We developed several supervised systems with the help of different classification algorithms and neural networks such as Support Vector Machines, Decision Trees, and Feed Forward Neural Networks. Our experiments reveal that features like timbre, rhythm, and intensity are associated with enhanced classification accuracy. Overall, the results were found satisfactory and Feed Forward Neural Networks based system achieved the maximum F-measure of 0.725 based on 10-fold cross validation.