Deep Learning versus Machine Learning for Music Genre Classification: A Dual-Dataset Evaluation

Music genre classification is a key challenge in Music Information Retrieval (MIR) due to the complex temporal and spectral characteristics of audio signals. This study provides a comparative evaluation of traditional machine learning (ML) algorithms-including Support Vector Machines (SVM), kNearest Neighbors (k-NN), and Logistic Regression-and modern deep learning approaches, specifically Convolutional Neural Networks (CNNs) and a hybrid CNN + BiLSTM architecture. Experiments were conducted on the GTZAN and FMA-Small datasets using both raw audio waveforms and Mel spectrograms as input features. The results indicate that deep learning models consistently outperform traditional methods across all evaluation metrics. On the GTZAN dataset, both the CNN and the CNN + BiLSTM architectures achieved an accuracy of 91%, while on the FMA-Small dataset, the CNN + BiLSTM model attained the highest accuracy of 93%. These findings demonstrate the effectiveness of deep learning in capturing both spatial and temporal dependencies in audio data and offer practical insights into model selection based on performance, scalability, and data characteristics.