Data augmentation by modifying the frequency characteristics of sounds.

Abstract

This paper shows the usefulness of the data augmentation method that modifies the frequency characteristics of sounds in the problem of detecting wood broken sounds occurring in large earthquakes. In this study, a CNN model is used to classify the wood broken sounds and the other environmental sounds. The CNN model is trained with the spectrograms that are converted from sound of 1 second. First, it is investigated whether if the wood broken sounds are correctly classified or not even though the recording devices or environments are different between the training and the testing. The result shows that it is difficult to classify the wood broken sounds recorded with different devices. Next, the frequency characteristics of sounds are modified to simulate the sounds recorded with other types of devices. As a result, it is confirmed that the accuracy of classifying the wood broken sounds is improved by training the simulated sounds.