Swaragram based Residual Neural Architecture for Raag Identification in Indian Classical Music

Raag is the root of Indian Classical Music (ICM) in its arrangement, performance, improvisation, and organization. Applications of automatic raag identification include recommendation and indexing of ICM. This paper establishes a residual neural network (ResNets) approach that can differentiate a Raag from the swaragram representation of a song instead of previous methods on this task, which heavily depend on shallow machine learning. ResNets are very deep neural architectures that deal with the vanishing gradient issue and are usually used in the visual domain. Experiments were conducted on two standard datasets (Carnatic and Hindustani) with 40 and 30 Raags. The model achieved 79.9 % for the Carnatic and 90.4 % for the Hindustani dataset using the Swaragram. We have tested the model with many features and used the best feature to test against the state-of-the-art.