Deep Neural Network Simulates Tempos Estimated by Five Wind Instrumentalists from Playing Sheet Music

Abstract

Sound is generated by Kansei information processing performed by an instrumentalist playing a musical score. To construct datasets for Kansei information processing of musical scores, we have performed subjects’ experiments using melodies extracted from some sets of music scores. All Japan Band Contests have a lot of pieces of music, and they are all selected as the required pieces, so we extracted the melodies from them and made a set of music scores for the subject experiment. The results of five subjects’ experiments show that the subjects’ estimated tempo differs. Furthermore, the tempo estimated by the subject from the score was defined as a 2-class classification problem of fast or slow tempo and simulated by a deep neural network. The recognition rate of the training data was more than 99.8% for all subjects’ datasets, but the recognition rate of the evaluation data varied from 90.3% to 77.4%.