Music inpainting is the task of generating the missing parts of a musical piece. This technique has been applied to the music editing and restoration of old scores. In this study, we focus on monophonic music inpainting using the symbolic representation, which encodes musical elements such as notes, harmony, and rhythm into discrete symbols. Previous studies have proposed many methods of music inpainting using deep learning, which have improved the performance of the inpainting. However, the conventional models cannot complement music considering the musical pattern in the surrounding segment when the models are trained on limited data, such as music by deceased composers or music from minor genres. In this study, we focus on InpaintNet, a music inpainting model, and investigate the use of a pre-trained InpaintNet model, which is trained on a relatively large dataset. Then, we finetune whole or part of the InpaintNet model on small dataset of the target domain. We conducted a comparative evaluation of InpaintNet trained using multiple approaches under various training data sizes. As a result, using the pre-trained model, it was possible to obtain stable output even with limited training data.
Naemura et al. (Wed,) studied this question.