A Multimodal Approach to Bengali Music Genre Classification Using 1D CNN and ResNet-50

Music genre classification is a crucial task in Music Information Retrieval (MIR), where accurate categorization of music into genres is essential for enhancing music recommendation systems, organizing large music libraries, and facilitating content discovery on streaming platforms. This study addresses the specific challenges of Bengali music genre classification, which include the diversity of musical styles and the limited availability of annotated datasets. We present a multimodal approach that combines numerical features with Mel-spectrogram images using a 1D Convolutional Neural Network (CNN) and a ResNet-50 based transfer learning model. A comprehensive dataset of 720 Bengali songs across six genres was compiled for this research, and a series of classification models were developed and evaluated. The proposed multimodal model demonstrated superior performance, achieving a test accuracy of 84.72%, a significant improvement over individual models. Additionally, the study discusses the pitfalls of certain practices in audio classification, such as data augmentation, which can lead to misleading performance metrics. The findings highlight the effectiveness of a multimodal approach in handling the complexities of Bengali music genres and suggest pathways for future research, including dataset expansion and the incorporation of additional modalities such as lyrics.