Abstract:

In music source separation, a standard data augmentation technique involves creating new training examples by randomly combining instrument stems from different songs. However, these randomly mixed samples lack the natural coherence of real music, as their stems do not share a consistent beat or tonality, often resulting in a cacophony. Despite this apparent distribution shift, random mixing has been widely adopted due to its effectiveness. In this work, we investigate why random mixing improves performance when training a state-of-the-art music source separation model and analyze the factors that cause performance gains to plateau despite the theoretically limitless number of possible combinations. We further explore the impact of beat and tonality mismatches on separation performance. Beyond analyzing random mixing, we introduce ways to further enhance its effectiveness. First, we explore a multi-segment sampling strategy that increases the diversity of training examples by selecting multiple segments for the target source. Second, we incorporate a digital parametric equalizer, a fundamental tool in music production, to maximize the timbral diversity of random mixes. Our experiments demonstrate that a model trained with only 100 songs from the MUSDB18-HQ dataset, combined with our proposed methods, achieves competitive performance to a BS-RNN model trained with 1,750 additional songs