Exploring the Role of Entropy in Music Classification

Music classification and analysis has been a very active area of study for quite a long time. While most of the research revolves around Convolutional Neural Networks (CNNs), we have employed information theoretic metrics to classify the songs. Abrupt amplitude changes in the music from time to time allow us to classify the music. We have used Shannon’s entropy and Kullback-Leibler divergence to capture and quantify the changes in amplitude and based on this we have classified the songs into three categories, (1) melody (2) neutral, and (3) rock. To classify a song, we divide the music/song into small frames and apply the statistical and information theoretical methods to each frame. These results are then combined to obtain a single value representing the entire song, which is used for decision-making. Regardless of the language and instruments used, melody songs generally exhibit lower entropy values compared to rock songs. Our study includes songs from various languages such as English, French, Hindi, Japanese, Malayalam, Spanish, Tamil, and Telugu. Using the entropy values on the random forest model, we achieved a maximum accuracy of 82% in song classification. This study can enable a new way of classifying music without too much reliance on computational capacity.