A Transfer Learning Strategy for Owl Sound Classification by Using Image Classification Model with Audio Spectrogram

Abstract

This paper presents an improved approach to train models that can be used to accurately predicting animal presence based on its sound with a limited dataset. Currently, deep-learned models dominate the state-of-the-art methods for audio classification tasks for their predictive capabilities. However, an immense amount of data is needed to build an accurate deep learned classifier. Such immensity of data is usually hard to be satisfied on endemic or endangered animals. For example, collecting an Indonesian scops owl audio dataset for our experiment in an adequate amount is insatiable, thus may reduce the predictive capability of a deep-learned model. To overcome such an issue, we propose a transfer learning strategy that alleviates overfitting in a deep model and a way to maximize the use of datasets by extracting two acoustic features: Mel's spectrogram and Mel Frequency Cepstral Coefficient (MFCC) from each data point. In this study, we employ a dual-input scalable Convolutional Neural Network (CNN) derived from EfficientNet [1] which utilizes and learn from both acoustic features. Our experimental pretrained dual-input network achieves 99.27% mAP on our testing data accuracy whereas a trainedfrom-scratch Resnet-50 model used as the baseline model achieves 99% mAP on the same testing set.