Swara Histogram Based Structural Analysis And Identification Of Indian Classical Ragas

This work is an attempt towards robust automated analysis of Indian classical ragas through machine learning and signal processing tools and techniques. Indian classical music has a definite heirarchical structure where macro level concepts like thaats and raga are defined in terms of micro entities like swaras and shrutis. Swaras or notes in Indian music are defined only in terms of their relation to one another (akin to the movable do-re-mi-fa system), and an inference must be made from patterns of sounds, rather than their absolute frequency structure. We have developed methods to perform scale-independent raga identification using a random forest classifier on swara histograms and achieved state-of-the-art results for the same. The approach is robust as it directly works on partly noisy raga recordings from Youtube videos without knowledge of the scale used, whereas previous work in this direction often use audios generated in a controlled environment with the desired scale. The current work demonstrates the approach for 8 ragas namely Darbari, Khamaj, Malhar, Sohini, Bahar, Basant, Bhairavi and Yaman and we have achieved an average identification accuracy of 94.28% through the framework.