抄録
Automated piano player systems enable highly precise keystrokes and pedal operations. However, a direct translation of score data into performance often results in a mechanical and inexpressive sound, failing to capture the nuanced dynamics of a human pianist's interpretation. This discrepancy arises because pianists uniquely determine various performance parameters, such as note loudness (Velo), the time between notes (Step), and individual note durations (Gate). While deep learning has been utilized in prior research to predict these parameters, the accuracy of Velo prediction, in particular, has remained a significant challenge. To address this limitation, this paper proposes a novel deep learning system specifically designed to enhance the accuracy of Velo prediction by integrating two distinct neural networks. Furthermore, experiments conducted with our system demonstrate improved prediction accuracy compared to previous studies.
https://ror.org/02278tr80 
