Overview
Music Genre Classification holds significant importance in the context of music content distribution platforms. In this study, we present BanglaBeats, a comprehensive dataset of Bengali songs. We also propose a CNN-based approach that surpasses previous CNN-based Bangla music genre classification approaches in terms of accuracy. We did not stop here. We explored the possibility of transformer-based approaches for Bengali music genre classification tasks. We fine-tuned DistiHubert and Wav2-Vec2-Base-960h models. Our results reveal that further optimization techniques can enhance this task.
Comprehensive Dataset
At the heart of our study lies a meticulously curated dataset of Bangla songs. Comprising 1617 data samples spanning across 8 distinct genres, it forms the foundation upon which our research thrives.
Innovative Approaches
Our Convolutional Neural Network (CNN)-based approach emerged as a standout performer, achieving a remarkable accuracy rate of 88%. This achievement not only surpassed existing methods but also demonstrated the potential of CNNs in music genre classification.
Fine-Tuned Transformers
We didn’t stop there. We delved into the realm of transformer-based models, meticulously fine-tuning DistilHubert and Wav2Vec2-Base-960h. Our rigorous efforts yielded impressive results with 83.36% and 84.94% accuracy, respectively. These outcomes showcase the versatility of modern deep learning models in deciphering complex Bengali musical patterns.
Future Horizons
Our journey continues as we explore advanced optimization techniques, aiming to push the boundaries of performance even further. We are trying to refine our approach, expanding our dataset, and contributing to the evolving landscape of Bengali music genre classification.