2026 Volume 34 Pages 974-986
The aim of this research is to improve the accuracy of difference-detection of piano performance mistakes in our proposed method based on anomaly detection. In previous research, we converted the performance data of learners and teachers into mel-spectrogram images and used an anomaly detection technique to evaluate the skillfulness of a learner's performance from the similarity of the images. While a high level of difference-detection accuracy was observed for the treble register in particular, there were problems concerning the bass register. Therefore, in this research, in the process of converting teacher and learner data into images, we applied Constant-Q Transform, in which window length can be adjusted in accordance with frequency to enable frequency resolution that is suited to the pitch. We conducted an experiment, involving four phrases, using the mel-spectrogram, used in our previous research, and the proposed method (OR condition of mel-spectrogram and Constant-Q Transform), and evaluated the detection rate. As a result of evaluation using RER (Relative Error Reduction), an index of the extent to which the rate has improved from the baseline, the average RER was 0.26 for the treble register, 0.57 for the bass register, and 0.50 overall, showing improvement centred on the bass register. These results demonstrated that in places where overtones become mixed in on a mel-spectrogram, such as where there are bass notes and chords, the detection rate can be improved by applying Constant-Q Transform.
