Abstract
In the genre of "rhythm action games," many games emphasize the experience of performing a musical piece by producing sounds according to the player's action. In this paper, automatic estimation of "key-sound" was performed for the purpose of automatic generation of the game data of rhythm action games. A key-sound is the sound object extracted from the music and played according to the player's action. We applied machine learning to estimate whether or not a specific sound object at a certain time should be a key sound, using the acoustic features and time series information of sound events. As a result, it was found that the accuracy of about 90% was achieved when the training and estimation were conducted using the data of the same musical piece. Even when the training and estimation were conducted using different musical pieces, the accuracy was higher than 60%. The analysis results on the acoustic characteristics of the key-sounds and their time series are expected to form the basis of automatic generation of the game data of rhythm action games.
