Evaluation of the impact of the Multi-Head Attention algorithm in Music Source Separation

This work focuses on the evaluation of the impact of the Multi-Head Attention algorithm in the field of Music Source Separation. In particular, our objective is to determine its potential to outperform the U-Net architecture often employed in state-of-the-art (SOTA) models. Additional primary goals include examining the repercussions of integrating Self-Supervised features into the pipeline and assessing the efficacy of the Attention mechanism for phase estimation. Notably, when utilizing the magnitude spectrogram as input, our model demonstrated promising outcomes, especially when using an increased volume of training data. The incorporation of Self-Supervised features into the model's architecture proved to be effective only when all layer representations are combined into a weighted sum. Blindly concatenating the last layer appeared to be less beneficial to the model's performance. Other findings in this thesis include confirming the utility of the SAD step in the preprocessing pipeline and analyzing the depth of the model, emphasizing once again that Music Source Separation (MSS) models encounter difficulties when the depth is too high. Lastly, it is observed that the Attention mechanism alone is insufficient for accurate phase estimation, a complex task not well suited for the chosen algorithm.