Bleeding Matrices: A New Tool for Music Source Separation Analysis and Evaluation

Abstract

Signal-to-distortion, signal-to-interference, and signal-to-artifacts ratios are the most common objective metrics for evaluating music source separation systems. The signal-to-interference ratio measures how much of the error in a separated stem comes from other sources not having been completely removed, but does not provide information about which sources are those. Inspired by the confusion matrices used to analyze and evaluate classification systems, we propose the use of bleeding matrices to visualize how much of the interference in every separated stem comes from every interferent source.