Swaram Extraction from Veena Tunes using Deep Learning Algorithms

Abstract

Music is one of the fast-growing industries in today's world and faces many difficulties. Indian Classical music, too, is different from other western music patterns and difficult to learn or identify. A lot of work has been done before using traditional approaches and expertise to identify Swaram(notes) with the help of standard features such as arohana, avarohana, pakad, gamak, vaadi, savandi etc. but with the help of new features such as chromagrams. Audio information retrieval (AIR) is a field with potential applications in automatic annotation, music recommendation, as well as music tutoring and accuracy verification systems. Extracting the Swaram (notes), or melodic style, of improvisational Classical music is a challenging problem in AIR due to the music's melodic variation and inconsistent temporal spacing. In this project, we propose a deep learning-based approach to Swaram (notes) recognition. Deep learning system is proposed for extracting information from audio data with temporal variation. Our method makes effective use of long short-term memory based recurrent neural networks to efficiently pre- possess and learn temporal sequences in music data (LSTM-RNN). Swaram identification is a sequence classification task in which each Swaram is treated as a class and the notes produced by prevailing melody estimation are treated as words. We train and test the network on smaller sequences sampled from the original audio while the final inference is performed on the audio as a whole.