The paper summarizes the music demixing (MDX) track of the Sound Demixing Challenge 2023 (SDX'23). It highlights the importance of robust music source separation (MSS) in the presence of errors and inconsistencies in the training data.
The authors first discuss the impact of label errors and bleeding (i.e., signal from one instrument bleeding into the recording of another) in the training data on the convergence and performance of MSS models. They then formalize these two types of errors and introduce two new datasets, SDXDB23_LabelNoise and SDXDB23_Bleeding, to simulate such errors.
The paper describes the methods that achieved the highest scores in the competition, including an iterative refinement baseline that uses the trained model to improve the quality of the training data. The authors also present a direct comparison with the previous edition of the challenge, showing an improvement of over 1.6dB in signal-to-distortion ratio (SDR) for the best performing system.
Additionally, the authors report the results of a listening test conducted with renowned producers and musicians to study the perceptual quality of the top systems. Finally, they provide insights into the organization of the competition and their prospects for future editions.
翻譯成其他語言
從原文內容
arxiv.org
深入探究