Automatic Orchestration of Piano Scores for Wind Bands with User-Specified Instrumentation

We present a deep learning method for generating wind band scores with user-specified instrumentation from piano scores. The difficulty in curating large-scale pair data with accurately aligned wind band and piano scores poses two major challenges: (i) efficient preparation of training data and (ii) effective learning of orchestration rules, particularly for infrequently used instruments. To address these problems, we propose using an automatic piano arrangement method to generate pair data from existing wind band scores. Our method utilizes U-Net to assign notes in an input piano score to individual instrument parts, and we propose refined network architectures for efficient learning of characteristics of instrument parts in the wind band scores. We show that the method can generate partially playable scores that capture voicing rules and mutual relationships among instrument parts.