A Dataset of rhythmic pattern reproductions and baseline automatic assessment system

Abstract

This work presents a novel dataset comprised of audio and jury evaluations for rhythmic pattern reproduction performances by students applying for a conservatory. Data was collected in-loco during entrance exams where students were asked to imitate a set of rhythmic patterns played by teachers. In addition to the pass or fail grades provided by the members of the jury during the exam sessions, a subset of the data was also evaluated by external annotators on a 4-level scale. A baseline automatic assessment system is presented to demonstrate the usefulness of the dataset. Preliminary results deliver an accuracy of 76% for a simple pass/fail logistic regression classifier and a mean average error of 0.59 for a linear regression grade estimator. The implementation is also made publicly available to serve as baseline for alternative assessments systems that may leverage the dataset.