第５部門①

[2E01-05-03] Study on Data Augmentation Method for Detecting a Wood Broken Sound by using CNN

*Mayuko ONO¹, Tsukasa TAKEMORI¹, Masayuki SAEKI¹ (1. Tokyo University of Science)

Keywords:Convolutional Neural Network, wood broken sound, data augmentation

The objective of this study is to find out the effective data augmentation method for detecting a wood broken sound by using CNN. In this study, the time stretching of sound and the noise addition are investigated as a data augmentation method. The simulation results show that those methods are very effective in improving the detection accuracy in case that the test and learning data are recorded in the same environment. On the other hand, the use of sounds recorded in various environments is needed to improve the accuracy in case of testing sounds recorded in different environment.