Evaluating Music Embeddings for Song Genre Classification

Music Genre Classification is an important application in the field of Music Information Retrieval, and it has got many commercial applications too. In this work, we perform various experiments to investigate the most appropriate music embeddings for this task. Three classification algorithms are implemented on two datasets (GTZAN and custom Hindi songs dataset). Experimental results show that stand-alone features are outperformed by certain fused features in the short-term spectral domain. Moreover, pre-trained features extracted from the deep YamNet model surpass the performance of conventional features, demonstrating the applicability of deep models to extract hidden patterns from the dataset for performing any related underlying task.