Automatic evaluation of music students

Abstract

Music Schools would greatly benefit from a tool that would reduce the time teachers spend evaluating students while those teachers would be confident that evaluation is still a reliable and perhaps more transparent process. We propose a method for evaluating students by asking them to play a specific piece of music, then we transform the audio-signal into a sequence of feature vectors. We align this sequence to a reference to determine how well the student played. We found that Dynamic Time Warping works better than Levenshtein distance or the Longest Common Subsequence distance to determine the similarity between the music played by the student and the reference. We tested our system with 28 musical performances played by music students and compare the grades determined by our system with the actual grades given to them by their teacher getting very encouraging results.