This paper presents a deep learning method for recognizing Ragas, a significant element in Indian Classical Music (ICM). The method uses Long Short Term Memory based Recurrent Neural Networks (LSTM-RNN) to learn temporal sequences in music data. The model was trained on smaller sequences taken from the original audio, with inference performed on the whole audio. It achieved an accuracy of 88.1% on the Comp Music Carnatic dataset, making it the state-of-the-art for Raga recognition. The method also enables sequence ranking, useful for retrieving related melodic patterns from a music database.
Publication date: 16 Feb 2024
Project Page: https://ismir2019.ewi.tudelft.nl/?q=node/102
Paper: https://arxiv.org/pdf/2402.10168
