"Hardness" as a semantic audio descriptor for music using automatic feature extraction

Abstract

The quality of "hardness" in music is an attribute that is most commonly associated with genres like metal or hard rock. However, other examples of music raise the question of whether there is a genre-independent general dimension of "hardness" that can be obtained from the signal automatically based on psychoacoustical features. In listening experiments 40 subjects were asked to rate 62 music excerpts according to their hardness. Using MATLAB toolboxes, a set of features covering spectral and temporal sound properties was obtained from the stimuli and investigated in terms of their correlation with the subjective ratings. By means of multiple linear regression analysis a model for musical hardness was constructed which shows a correlation of r = 0.86 with the experimental results. This proposes musical hardness as a useful high level descriptor for analysing collections of music. In ongoing experiments the fitness of this model is being further evaluated.