Modelling musical attributes to characterize two-track recordings with bass and drums
In this publication, we present a method to characterize two-track audio recordings (bass and drum instruments) based on musical attributes. These attributes are modelled using different regression algorithms. All regression models are trained based on score-based audio features computed from given scores and human annotations of the attributes. We compare five regression model configurations that predict values of different attributes. The regression models are trained based on manual annotations from 11 participants for a data-set of 70 double-track recordings. The average estimation errors within a cross-validation scenario are computed as evaluation measure. Models based on Partial Least Squares Regression (PLSR) with preceding Principal Component Analysis (PCA) and on Support Vector Regression (SVR) performed best.