COMPARISON BETWEEN NN AND CNN MODELS DEVELOPED FOR DETECTING WOOD BREAKING SOUNDS

Abstract

This paper shows the usefulness of CNN with Mel-spectrogram in the detection of wood breaking sounds occurred in a large earthquake. In this research, NN and CNN models are used to detect wood breaking sounds. The NN model is trained by using the input data of MFCC which is calculated from various time series data of 3 seconds. On the other hand, the CNN model is trained by using the spectrogram or Mel-spectrogram which are also calculated from the same time series data. Then, the accuracies are compared to each other. As a result, the both methods show excellent accuracies. Furthermore, the other sounds which are synthesized by mixing a wood breaking sound with other environmental sounds are classified using the trained model. In this case, it is confirmed that the CNN with Mel-spectrogram can be more accurate classificatory than others.