Diffusion inspired training strategy for Source Separation in the frequency domain

The challenge of segregating individual musical components from a compound auditory signal is a longstanding issue in the realm of audio processing. Classical methodologies have provided quite interesting solutions but presenting lim-itations in robustness. Drawing inspiration from the recent success and popularity of diffusion models, this work hypothesizes that such training strategy, combined with a frequency-domain-based implementation, can offer substantial improvements in the task of music source separation. The leading-edge implementations of this topic hinge on conventional architectures such as U-net or Wave-U-Net and training approaches with little novelty. This means that the few improvements that can be achieved depend on the amount of data available and the depth of the neural net-work. This thesis approaches the problem from a different perspective, proposing a novel training strategy based on the nature of diffusion models. To this end, we will first approach the creation from scratch of a music source separation model with different variants and then apply the novel training strategy. The significance of this research lies not just in advancing the technical capabilities of music source separation, but also in the potential applications in areas such as music remixing, restoration, or automatic music transcription. This thesis, therefore, serves as a vanguard, encouraging further exploration and refinement in the domain of music source separation applying a diffusion-inspired training strategy.