Automatic Piano Fingering Estimation Using Recurrent Neural Networks

Abstract

Deciding piano fingerings is an essential skill for all piano players regardless of their expertise. Traditionally, pianists and piano educators first need to analyze musical scores, then they manually label the fingerings on the scores; however, this process is time-consuming and inefficient. This paper proposes a novel automatic piano fingerings estimating method by utilizing Bidirectional Long Short-term Memory (BI-LSTM) networks — a special type of Recurrent Neural Networks (RNNs). This is one of the first studies to explore the possibilities of applying deep learning to estimate piano fingerings. Together with the new method, a novel input representation is designed to capture the relations between surrounding notes. Furthermore, in addition to directly comparing the estimations with the ground-truth, this paper proposes a novel evaluation metric to assess the playability of the estimated fingerings. The results illustrate the effectiveness of the proposed method that generates playable and accurate estimated fingerings.