Evaluation of blind source separation algorithms applied to music signals

We present an evaluation of single-channel blind audio source separation (BASS) algorithms for different musical instrument combinations, based on statistical approach. So-called primitive BASS algorithms were chosen, because of their low computational latency and similarity to human hearing principles. The most common musical instruments (drums, guitars, vocals) were combined with each other, and segregate subsequently by 4 different blind audio source separation techniques. 25-second samples from 30 songs of various genres were used to preserve natural diversity. The main difference in our approach lies in the evaluation of average deviations from ground truth, instead of evaluating absolute values of signal-to-distortion ratio.