Score-informed transcription for automatic piano tutoring

Abstract

In this paper, a score-informed transcription method for automatic piano tutoring is proposed. The method takes as input a recording made by a student which may contain mistakes, along with a reference score. The recording and the aligned synthesized score are automatically transcribed using the non-negative matrix factorization algorithm for multi-pitch estimation and hidden Markov models for note tracking. By comparing the two transcribed recordings, common errors occurring in transcription algorithms such as extra octave notes can be suppressed. The result is a piano-roll description which shows the mistakes made by the student along with the correctly played notes. Evaluation was performed on six pieces recorded using a Disklavier piano, using both manually-aligned and automatically-aligned scores as an input. Results comparing the system output with ground-truth annotation of the original recording reach a weighted F-measure of 93%, indicating that the proposed method can successfully analyze the student's performance.