Archive
IJMSTA - Vol. 7 - Issue 1 - January 2025
ISSN 2612-2146
Pages: 11
RagaGEN: Fine-Tuning MusicGen Transformer Models for the Generation of Indian Classical Raga Melodies
Authors: Devananda Sreekanth, Sreekanth Gopi, and Femi Ojo
Categories: Journal
Abstract - The urgency to preserve and regenerate Indian classical music stems from a decline in its transmission and appreciation, especially among younger generations. Once rooted in oral pedagogy and vedic sacred texts, this musical tradition not only represents cultural identity but also holds therapeutic promise for addressing stress, anxiety, and emotional disconnection in modern life. In this context, the integration of AI-driven techniques for music generation holds immense potential to bridge the divide between creativity and technology, thereby facilitating the preservation of cultural heritage and enriching music education. In this study, we investigate the integration of traditional Indian music with modern AI techniques by fine-tuning MusicGen for the generation of Raga melodies. Our literature review identified a notable gap in utilizing autoregressive language models like MusicGen for Raga melody production. MusicGen, known for its ability to generate music from textual prompts, was fine-tuned to learn and produce melodies characteristic of the 'Jog' Raga. Due to the scarcity of the Jog Raga melody music dataset, we developed a dataset comprising 20 singing voices in Jog Raga, converted it into MIDI format, and subsequently transformed it into WAV files for fine-tuning with MusicGen. The methodology encompassed a diligent fine-tuning process for the MusicGen model, involving dataset creation, pre-processing, prompt engineering, and extensive training through Dora to effectively generate music that captures the nuanced expressions of Jog Raga. The results demonstrated that the fine-tuned MusicGen model successfully produced compositions with 96% adherence to the Jog Raga's structures, effectively integrating key traditional elements and enhancing the authenticity of generated melodies. From the AI-generated tunes, we composed and published an instrumental track with violin, cello, santoor , and piano to showcase the potential usage of the generated melodies in professional music production. Moving forward, we would further train the model on a larger dataset of different Ragas and their characteristics, and deploy "RagaGEN"-our enhanced model capable of generating authentic Raga melodies with additional features and refinements.
Keywords: AI music generation, MusicGen, Raga music, Jog Raga, Traditional Indian music
Download paper
