Automatic evaluation of music students from MIDI data

A virtual music teacher just as a human music teacher needs to evaluate his students to decide if they should go to the next lesson or not. In this work the student is evaluated from symbolic data, specifically from the information recorded in MIDI files which are normally produced by modern instruments such as music keyboards. Our approach for evaluating a music student consists of converting the MIDI file produced by the student to a long string with the sequence of music notes and durations. This string is compared to another long string that corresponds to a “correct” version of the same music which has been generated by a trained musician. The more similar the string generated by the student to the string generated by the trained musician is the higher the grade. In our experiments 13 music students were evaluated and their grades compared to those assigned by a music teacher. According to our tests the Transposition invariant InDel distance proved to be the most adequate. The Branch and Bound algorithm was used to deal with the fact that this distance requires more time to be computed. The grades of the students are quite similar to those assigned by the teacher so the system can reliably and impartially be used to evaluate music students.