A Deep Autoencoder Approach To Bird Call Enhancement

Abstract

Due to their high performance, deep learning based approaches have attracted much attention in recent years. In this paper, we investigate deep autoencoder (DAE) based bird call enhancement. The DAE is trained utilizing layer-wise pretraining and then fine-tuned. Objective measures such as PSNR and log spectral distortion are used to compare the effectiveness of the DAE for enhancement. Furthermore, a deep neural network (DNN) based species classification system is utilized as an application to evaluate the effectiveness of the DAE. Our experiments indicate the effectiveness of the DAE based enhancement system, including the ability to provide more generalizable inputs for the DNN-based classifier. We also demonstrate the improvements obtained when bandpass filtering is performed as a preprocessing step for the DAE and the DNN.