Learned Features for the Assessment of Percussive Music Performances

摘要

The automatic assessment of (student) music performance involves the characterization of the audio recordings and the modeling of human judgments. To build a computational model that provides a reliable assessment, the system must take into account various aspects of a performance including technical correctness and aesthetic standards. While some progress has been made in recent years, the search for an effective feature representation remains open-ended. In this study, we explore the possibility of using learned features from sparse coding. Specifically, we investigate three sets of features, namely a baseline set, a set of designed features, and a feature set learned with sparse coding. In addition, we compare the impact of two different input representations on the effectiveness of the learned features. The evaluation is performed on a dataset of annotated recordings of students playing snare exercises. The results imply the general viability of feature learning in the context of automatic assessment of music performances.