Sinhala Music Genre Classification Using a Deep Learning Approach

Traditional approaches to genre classification have relied on user ratings, emotions, and genres, often utilizing Support Vector Machines (SVM) without leveraging deep learning methodologies for genre classification. To address this gap, this paper presents a comprehensive and novel investigation into the application of deep learning techniques based on Convolutional Neural Networks (CNNs) for Sinhala music genre classification using spectrogram images derived from Mel-frequency Cepstral Coefficients (MFCCs). The methodology involves data collection and preprocessing, model architecture design, training procedures, experimentation, and the presentation of results. A dataset comprising 325 Sinhala songs, with 65 songs representing each of five prominent genres (pop, hip-hop, baila, calypso, and classical), was compiled. From these audio files, 1300 MFCC spectrogram images were extracted for feature representation. Preprocessing steps included format conversion, audio segmentation, feature extraction, visualization, and normalization to prepare the data for input to the CNN model. The CNN architecture designed for genre classification consists of multiple convolutional and pooling layers followed by fully connected layers. Transfer learning with pre-trained models, including VGG-19, InceptionV3, EfficientNet, and DenseNet was explored. CNN and RNN hybrid models were explored by adding an LSTM layer followed by CNN layers. Hyperparameter tuning techniques, particularly Random Search, were employed to optimize model performance. Evaluation metrics such as accuracy, precision, recall, and F1-score were utilized to assess the effectiveness of the classification models. Results indicate promising performance, with the DenseNet model achieving the highest accuracy of 78.0%. However, challenges such as dataset limitations remain areas for improvement.