Data Augmentation for Improving the Performance of Raga (Music Genre) Classification Systems

The ability to identify the ragas (music genres) of Indian musical compositions is considered to be a rare talent or gift among musicians and music aficionados. It would take years of rigorous training and exposure to various ragas for elevating an ardent learner of music to such levels of expertise. Machine learning supported raga identification systems try to mimic this cognitive ability of an ace musician to recognize the patterns or signatures of ragas so that the ragas of music compositions can be identified. In this work, the audio samples obtained from the flute concerts of a legendary flutist were analyzed for recognizing the ragas. First a convolutional neural network (CNN) algorithm based raga classification system was developed with mel-spectrogram as input which yielded an accuracy of 52.4%. When this system was revamped with the techniques of data augmentation, the classification accuracy was substantially enhanced to 83%. This system can be made more versatile by incorporating more ragas in the corpus, by including signals from musical instruments other than flute and by incorporating signals from vocal concerts and also musical renderings set on different sruthis (tonic).