Automatic assessment of student rhythmic pattern imitation performances

In this study, we present an automatic performance assessment system for a student's rhythmic pattern imitation. The system compares the student's performance recording with a teacher's reference recording and assigns a grade between 1 and 4. For this task, we collected and publicly shared a data set that consists of recordings from real auditions for conservatory entrance exams in Turkey. The recordings are annotated/graded by three experts. The automatic assessment (grade assignment) task is considered as a regression problem and two approaches are tested and compared. The first approach applies classical regression methods to distance features. Distance features are computed by using several distance metrics to compute the distances between the performance and reference recording onset detection functions, binary onset vectors and onset times. The second approach applies metric learning using a Siamese neural network to directly learn the most efficient feature representations from the onset positions. We also investigate the effect of onset detection errors on automatic assessment performance. In the experiments, we observe that the assessment performance is significantly improved with an improvement in the onset detection accuracy. The best performance is achieved using the Siamese network with a mean absolute error (MAE) of 0.49. This performance is comparable to variation in expert annotations showing the potential of the system to be used in online music education. All the source codes and the data set (together with the annotations) are openly shared. We think that this completely reproducible work would potentially serve as a strong baseline for future studies on rhythmic pattern imitation tasks.