Source Separation for Carnatic Music
Indian classical Carnatic music comprises diverse sources such as mridangam, violin, ghatam, and multiple vocal lines (typically two or three). These sources, particularly in combination with the complex vocal styles, are significantly out-of-domain for existing state-of-the-art music source separation (MSS) systems. Consequently, developing specialized source separation models for Carnatic music can substantially enhance the performance of various downstream music information retrieval (MIR) tasks in the Indian music context.
To address this, a dedicated MSS model for Carnatic music has been developed using the Saraga dataset.
Approaches
- Fine-tuning several state-of-the-art MSS architectures, including Hybrid Transformer Demucs, Wave-U-Net, and TF U-Net.
- Designing a diffusion-based neural model tailored to the Carnatic music.
Trends in MSS
