"RaagaDhvani: A novel augmented multi-feature dataset: Advancing emotion recognition in Carnatic music with multimodal features and hybrid deep learning".

Carnatic music, a major form of South Indian classical tradition, offers rich potential for computational and musicological research. While datasets like CompMusic Carnatic Corpus, GTZAN, RAVDESS, IEMOCAP, EMODB, SAVEE, Saraga, and CMDB have advanced studies in raga recognition and pitch analysis, most lack emotion-based annotations essential for music emotion recognition (MER), affective computing, and therapeutic applications. This absence limits understanding of the nuanced emotional and structural dimensions of Carnatic music, underscoring the need for emotion-labelled datasets to support mental health and cross-cultural music research. To address this, the RaagaDhvani dataset has been created, focusing on culturally significant, emotion-driven vocal and flute renditions of eleven carefully selected Carnatic ragas, which are chosen for their psychological and affective associations. The dataset comprises 11 classical ragas, each recorded for approximately 5 min in high-quality audio. This approach preserves essential musical features, including gamakas, phrase structures, and tonal motifs, while generating sufficient training samples for deep learning models. The 448 ragas dataset is further enhanced with data augmentation, including pitch shifting, time-stretching, and noise addition. Later, to increase the validity and quality of Carnatic music, self self-curated dataset of 165 files are augmented with pitch shift and time stretch, resulting in 825 files to increase variability and improve model generalization. This dataset establishes a strong foundation for computational analysis of Carnatic music, supporting tasks like raga classification, emotion recognition, and multimodal retrieval. By integrating audio features with emotion annotations, it advances deep learning-based emotion prediction and fosters applications in psychology and music therapy.