Reduced Convolutional Recurrent Neural Network Using MFCC for Music Genre Classification on the GTZAN Dataset

This study presents a reduced Convolutional Recurrent Neural Network (CRNN) model for music genre classification, leveraging the GTZAN dataset and Mel-Frequency Cepstral Coefficient (MFCC) feature extraction. Unlike more complex architectures, this model simplifies the CRNN structure to three convolutional layers and two BiLSTM layers, maintaining competitive performance while reducing computational complexity. Key experimental parameters included learning rate tuning (0.1, 0.01, 0.001, and 0.0001) and dropout usage (30% before the BiLSTM layers) to mitigate overfitting. The best configuration, utilizing a learning rate of 0.001 and dropout, achieved an accuracy of 88.64%, outperforming more complex CRNN models by approximately 15%. These results underscore the potential of streamlined architectures in music information retrieval tasks, particularly for applications where computational resources are constrained. Future work will address overfitting issues and refine the dataset for enhanced model performance.