Automatic bird sound detection in long range field recordings using Wavelets & Mel filter bank features

The topic of bio-monitoring of fauna, especially that of birds, is an ongoing research topic. Although huge datasets of bird sound recordings are available, the classification of such sounds into bird & non bird sounds has been painstaking work, sometimes requiring manual processing. The goal of the IEEE Research Challenge in 2016 has been to address this concern and to help develop automatic algorithms for the detection of bird sounds [1]. An attempt has been made to compare and understand how wavelet based features perform against the current state of the art methods in audio processing in more detail. Two statistical/deep learning approaches namely Support Vector Machines and Convolutional Neural Networks have been used to compare the results. Since most of the works in audio signal processing have employed Fourier based methods, there has been a stagnation in the development and usage of newer/other features for experiments. This is one of the points that has been addressed in this work. The Receiver Operating Characteristic curve (ROC) has been employed to test the diagnostic ability of the audio features. Wavelets have performed consistently and at a level similar to the best performing Fourier based methods highlighting the possibility of using such features as a viable alternate for future audio processing experiments.