Can AI Read Sheet Music Like Human Beings?

In this paper, we propose a machine learning model to predict tempo using sheet music alone. The model estimates tempo in a manner analogous to how humans interpret affective and contextual information in sheet music. To construct the dataset for training, we invited wind instrumentalists to provide tempo annotations. After applying data augmentation techniques, the model was trained and evaluated through experiments. The results show that for the training data, 98.0% of the predicted tempo values deviated by less than 10 from the expected values. For the validation data, this percentage was approximately 62.1%. Notably, sheet music with significant tempo deviations is difficult to judge, even for human instrumentalists. Overall, the model demonstrated the ability to predict tempo from sheet music, achieving performance comparable to that of a beginner instrumentalist.