Abstract

The assessment of musical performances in, e.g., student competitions or auditions, is a largely subjective evaluation of a performer’s technical skills and expressivity. Objective descriptors extracted from the audio signal have been proposed for automatic performance assessment in such a context. Such descriptors represent different aspects of pitch, dynamics and timing of a performance and have been shown to be reasonably successful in modeling human assessments of student performances through regression. This study aims to identify the influence of individual descriptors on models of human assessment in 4 categories: musicality, note accuracy, rhythmic accuracy, and tone quality. To evaluate the influence of the individual descriptors, the descriptors highly correlated with the human assessments are identified. Subsequently, various subsets are chosen using different selection criteria and the adjusted R-squared metric is computed to evaluate the degree to which these subsets explain the variance in the assessments. In addition, sequential forward selection is performed to identify the most meaningful descriptors. The goal of this study is to gain insights into which objective descriptors contribute most to the human assessments as well as to identify a subset of well-performing descriptors. The results indicate that a small subset of the designed descriptors can perform at a similar accuracy as the full set of descriptors. Sequential forward selection shows how around 33% of the descriptors do not add new information to the linear regression models, pointing towards redundancy in the descriptors.