Raga Classification in Carnatic Music Using Audio Thumbnailing

The concept of raga is integral to Indian classical music theory and is inseparable from its practice. Hence, Raga classification is an important problem for Music Information Retrieval (MIR) Systems. However, the sheer number of ragas and the variations in their structure in Carnatic music make classification/identification of Ragas a difficult task. Several approaches have been proposed in research literature for Raga classification using various machine learning methods. However, all of them use entire songs for training the proposed models. As part of this paper, we propose a method to classify Raga using a representative structure of a song called audio thumbnail.We used parallel CNN - RNN model to classify the ragas. Further, to show the robustness of our approach, various experiments based on 16 different ragas were carried out and gave us the prediction accuracy of 83% on average.