On Investigating a Better Audio Representation for Mood Classification in Indian Popular Music

Developing a specialized music mood classifier for Indian popular music is essential due to the diverse musical traditions and languages in India. It enhances music recommendation systems and music therapy applications by accurately capturing Indian popular songs' emotional nuances. Considering the complexity of Indian popular music, the existing approaches in Western music have limitations. This investigation examines whether existing large audio models can improve the accuracy of music mood classification by generating more informative audio representations. Our dataset includes more than 338 songs from South Indian languages, ensuring diversity. The proposed architecture utilizes HuBERT and MERT representations of the audio. On comparing with the CNN-based baseline and a few related architectures, the proposed model performs better highlighting the potential of the music representation for mood classification in Indian popular music. Additionally, our analysis of pre-2000 songs revealed unique emotional tones that complicate predictions, emphasizing the stylistic differences between pre-2000 and post-2000 music.