Exploring Empirical Mode Decomposition for Music Genre Classification Using Deep Learning

Music streaming platforms have seen a surge in users. They store a huge catalogue of music in their databases, and it would be an extremely tedious and error-prone process to manually organize such a vast library into categories. This has motivated research aimed at automating this task efficiently and economically. Systems employed for this task require much information about the music signal. Empirical Mode Decomposition is a signal adaptive decomposition method used for obtaining characteristic information about the signals. We suggest a novel method of using Mel-Frequency Cepstral Coefficients extracted from the Intrinsic Mode Functions obtained after Empirical Mode Decomposition of the music signal. We have employed two Deep Learning models: Convolutional Neural Network and Convolutional Recurrent Neural Network, for the multi-class classification task. We have further compared the performance of the models on individual IMFs. We did not observe any improvement in the CRNN model. Further, we observed that the performance generally decreased with every successive IMF. The highest accuracy was obtained using the MFCCs extracted from the first IMF in the CNN model (91.27%).