Training Data Clustering for Key-Sound Estimation in Rhythm Action Games

Abstract

A rhythm action game is a genre of video game in which the player acts in accordance with visual symbols (charts) representing the rhythm of music. This work aims at generating a chart from the music automatically to realize easy development of rhythm action games. A key-sound is a sound object which is played in response to the player's operation. To generate a chart, it is necessary to classify the sound objects into key-sounds and the other sounds. This process can be done by machine learning, but the training data have a wide variety due to the individuality of creators, and the optimal chart is not unique. Accordingly, we divided the training data into clusters based on the classification tendency and prepared multiple models. Classification experiments using multiple models suggested the possibility of improvement, but further investigation is necessary for the selection of the optimum cluster.