An Improvement of Piano Fingering Recommendation System Using Neural Networks

Abstract

Using neural networks to predict piano fingering poses a challenge in recommending appropriate fingerings, with previous studies achieving an accuracy of around 60%. Therefore, in this study, we aim to improve accuracy by considering previous fingerings and keystroke timings as new parameters. Furthermore, to confirm the improvement in accuracy, we conducted comparative experiments between models such as RNN, FFNN, BiRNN, and the recently renowned Vision Transformer, known for its high accuracy. We compared experiments with and without the addition of new parameters. As a result, considering previous fingerings increased the accuracy by an average of 8%, while considering keystroke timings raised it by approximately 1%. Additionally, among the four types of neural networks, it was confirmed that BiRNN had the highest accuracy.