Analysing Scattering-Based Music Content Analysis Systems: Where's the Music?

Abstract

Music content analysis (MCA) systems built using scattering transform features are reported quite successful in the GTZAN benchmark music dataset. In this paper, we seek to answer why. We first analyse the feature extraction and classification components of scattering-based MCA systems. This guides us to perform intervention experiments on three factors: train/test partition, classifier and recording spectrum. The partition intervention shows a decrease in the amount of reproduced ground truth by the resulting systems. We then replace the learning algorithm with a binary decision tree, and identify the impact of specific feature dimensions. We finally alter the spectral content related to such dimensions, which reveals that these scattering-based systems exploit acoustic information below 20 Hz to reproduce GTZAN ground truth. The source code to reproduce our experiments is available online.