[ICASSP XAI-SA 2024] Why does music source separation benefit from cacophony?

Former MERL Intern Chang-Bin Jeon presents his paper titled "Why does music source separation benefit from cacophony?" for the IEEE ICASSP Satellite Workshop on Explainable Machine Learning for Speech and Audio (XAI-SA), held in Seoul (South Korea) Apr 15 2024. The paper was co-authored with MERL researchers Gordon Wichern, François G. Germain and Jonathan Le Roux.
Paper: merl.com/publications/TR2024-030
Abstract: In music source separation, a standard training data augmentation procedure is to create new training samples by randomly combining instrument stems from different songs. These random mixes have mismatched characteristics compared to real music, e.g., the different stems do not have consistent beat or tonality, resulting in a cacophony. In this work, we investigate why random mixing is effective when training a state-of-the-art music source separation model in spite of the apparent distribution shift it creates. Additionally, we examine why performance levels off despite potentially limitless combinations, and examine the sensitivity of music source separation performance to differences in beat and tonality of the instrumental sources in a mixture.