An Enhanced CNN Architecture for Music Genre Classification on the GTZAN Dataset
Joshua Wilson, Ansh Rao, Ayesha Taranum, Himanshu Gupta, Mohammed Nafishuddin · 2025
A principal objective within contemporary Music Information Retrieval (MIR) research is the development of automated systems for genre classification, especially due to the exponential proliferation of digital audio content on platforms such as streaming services, online radio, and algorithmically generated playlists. Manual annotation is no longer viable, thereby necessitating scalable and intelligent classification solutions. Music Genre Classification Using Convolutional Neural Networks presents a comprehensive examination of automatic music genre classification using deep learning frameworks, augmented by signal processing and traditional machine learning methodologies. The GTZAN genre collection, comprising 1,000 audio tracks each with a duration of 30 seconds, serves as the primary dataset. This benchmark includes ten balanced musical genres: blues, classical, country, disco, hip-hop, jazz, metal, pop, reggae, and rock. Feature extraction is performed using both time-domain and frequencydomain techniques. To address the challenges inherent in modeling complex, high-dimensional audio data, Music Genre Classification Using Convolutional Neural Networks proposes a specialized CNN architecture that utilizes log-mel spectrogram representations of the audio signal as two-dimensional input. Data augmentation techniques such as noise injection, pitch shifting, and time stretching are employed to improve model robustness and generalization across diverse musical content. The CNN model achieves an average classification accuracy of$\mathbf{9 5. 2 \%}$, demonstrating strong capability in learning genrespecific acoustic patterns. Analysis of the confusion matrix reveals classification challenges in genres with overlapping sonic characteristics, such as classical and jazz or rock and metal. Nevertheless, the high precision and recall across most categories affirm the effectiveness of CNN-based methods for music genre recognition and their applicability to large-scale music retrieval and recommendation systems.