MUSIB: musical score inpainting benchmark

Abstract Music inpainting is a sub-task of automated music generation that aims to infill incomplete musical pieces help musicians in their composition process. Many methods have been developed for this task. However, we observe tendency each method be evaluated using different datasets and metrics the papers where they are presented. This lack standardization hinders an adequate comparison these approaches. To tackle problems, present MUSIB, new benchmark score with standardized conditions evaluation reproducibility. MUSIB evaluates four models: Variable Length Piano Infilling (VLI), InpaintNet, SketchNet, AnticipationRNN, over two commonly used datasets: JSB Chorales IrishFolkSong. We also compile, extend, propose adequately quantify note attributes such as pitch rhythm Note Metrics , but higher-level properties introduction Divergence which operate by comparing distance between distributions features. Our shows VLI, model based on Transformer architecture, best performer larger dataset, while VAE-based models surpass Transformer-based relatively small dataset. With aim at inspiring community towards better reproducibility research, setting example strongly founded comparisons among SOTA methods.