Music Source Components Isolation using MultiHead Transformers and Encoders

A novel approach for isolating music source components using a combination of multi-head transformers and encoders is proposed in this paper. The pursuit of delineating individual elements within intricate audio signals, such as drums, piano, vocals, and assorted instruments, has become increasingly paramount in contemporary audio processing research as well it is an important part for music producers. Traditional methods for music source separation have often relied on handcrafted features and signal processing techniques, which may struggle to capture the intricate dependencies and nuances present in musical audio. In this study, we propose a data-driven approach that leverages the power of multi-head transformers and encoders to automatically learn and extract meaningful representations of different music source components directly from raw audio waveforms. Our model is capable of efficiently separating and isolating various musical elements, enabling enhanced analysis, manipulation, and understanding of music content. We demonstrate the effectiveness of our approach through extensive experiments on diverse music datasets, showcasing superior performance compared to existing methods. Our findings emphasize the potential applications and future directions for music source component isolation in domains such as music production, audio engineering, and music information retrieval.