<p>Cross-Cultural Music Genre Classification With Self-Supervised Music Transformers: A MERT-Based Transfer Learning Study on Congolese and International Genres</p>

Automatic music genre classification has progressed rapidly with deep representation learning, yet most benchmark-driven studies remain concentrated on Western musical repertoires. This imbalance limits the evidence available for culturally diverse music information retrieval systems and may lead to models whose reported accuracy is not representative of under-studied musical traditions. This study evaluates a transfer learning pipeline for cross-cultural genre classification by combining 10 international genres from the GTZAN corpus with five Congolese genres: rumba, soukous, sebene, mutuashi, and omounde. A balanced 15-class corpus of 1,500 thirty-second audio clips is processed using a standardized audio pipeline and represented through fixed 1,024-dimensional embeddings extracted from a pretrained Music Embedding Representation Transformer (MERT). Four downstream classifiers are compared: a radial-basis-function support vector machine, random forest, multilayer perceptron, and a compact fully connected deep neural network. The multilayer perceptron obtains the best within-corpus performance, reaching 93.00% accuracy and 93.02% F1-score, followed by the support vector machine and deep neural network at 91.00% accuracy. Confusion matrix and t-distributed stochastic neighbour embedding analyses show that several Congolese genres form compact and discriminative regions in the embedding space, while most residual errors occur between musically adjacent international genres such as blues, country, and rock. The findings indicate that self-supervised music representations can support culturally inclusive genre classification under limited labelled data, provided that dataset provenance, leakage control, and external validation are handled rigorously.