Music Genre Classification Using Contrastive Dissimilarity

In the digital age, streaming platforms have revolutionized how we access and interact with music, highlighting the need for more intuitive ways to organize and categorize our ever-growing music collections. The challenge lies in effectively classifying tracks into similar genres and styles to enhance user experience through improved music discovery and recommendation. In this context, machine learning stands out as a powerful tool. Traditional research in the field focuses on the auditory characteristics of music, such as timbre and rhythm. Nevertheless, the incorporation of spectrogram analysis introduces a richer layer of data representation, capturing the intricate musical textures that distinguish genres. This study proposes a novel approach to music genre classification, leveraging classic machine learning algorithms and the recently proposed contrastive dissimilarity method. Our methodology, which involves a detailed examination of spectrograms and the use of conventional feature extraction methods such as Local Binary Patterns (LBP), Local Phase Quantization (LPQ), Binarized Statistical Image Features (BSIF), and Oriented Basic Image Features (OBIF), combined with deep neural embeddings estimated using the contrastive dissimilarity method, offers a more comprehensive and accurate way to classify music genres. Our comparative analysis, conducted on three benchmark music genre datasets - GTZAN, Latin Music Database, and ISMIR 2004 - demonstrates promising results that approach the performance of current state-of-the-art methods.