Data mining and game sounds classification prerequisite to find a compact but effective set of features in the overall problem-solving process. As a preprocessing step of data mining, feature selection has tuned to be very efficient in reducing its dimensionality and removing irrelevant data at hand. In this paper we cast a feature selection problem on rough set theory and a conditional entropy in information theory and present an empirical study on feature analysis for classical instrument classification. An new definition of a significance of each feature using rough set theory based on rough entropy is proposed. Our results suggest that further feature analysis research is necessary in order to optimize feature selection and achieve better results for the musical instrument sound classification problem through Weka’s classifiers. The results show that the performance of the best 17 selected features among 37 features has 3.601 compared to 2.332 in standard deviation and 94.667 compared to 96.935 in average with four classifiers.