An efficient audio based performance evaluation system for computer assisted piano learning
In this paper, we propose an audio based piano performance evaluation system for piano learning, aiming at giving objective feedbacks to the piano beginners so that their self-practicing could be more efficient. We target to build a system which could evaluate the performance of an input recorded piano audio signal like an expert. First, we extract the Constant-Q transform (CQT) spectrum from audio signals as the feature sequences. Then, we employ Dynamic Time Warping(DTW) algorithm on the CQT feature sequences to align the input with the template. On top of the aligned feature sequence pair, we extract multiple global features that relating with the similarities and rhythms. Finally, we apply linear regression upon these global features with the training data to estimate the expert generated performance scores for test audio. Experimental results show that our system achieves 2.89 average absolute score error and 0.77 average correlation coefficient on our in-house collected dataset.
