BanglaBeats: A Comprehensive Dataset of Bengali Songs for Music Genre Classification Tasks

Music genre classification holds significant importance in the landscape of music content distribution. The accurate classification of music genres can enhance the overall user experience by employing various services in audio distribution and streaming platforms. Traditional deep-learning approaches like Convolutional Neural Networks have shown promising results in music genre classification tasks. Pre-trained models such as DistilHubert and Wav2Vec2-Base-960h also show very good results. However, there has been a lack of comprehensive Bengali music datasets for genre classification tasks. Existing datasets lack fusion-imposed recent audio samples which are hard to classify into genres. Previous Bengali music genre classification approaches also did not explore transformer-based pre-trained models. In this study, we present BanglaBeats, a Bengali music dataset containing 8 genres and 1617 music data samples. This dataset consists fusion styled audio samples, recent as well as old samples to properly represent the realm of Bengali music heritage. Along with the dataset, we also developed a CNN model tailored for Bengali music genre classification, which achieved exceptional performance, reaching a test accuracy of 88%. This surpassed all other existing CNN models in this domain. Additionally, we explored the effectiveness of pre-trained models, specifically DistilHubert and Wav2Vec2-Base-960h, and obtained impressive test accuracies of 83.36% and 84.94%, respectively. These findings emphasize the efficacy of our proposed model and the potential of transformer-based models for Bengali music genre classification.