Understanding Variable Performance on Deep MIL Framework for the Acoustic Detection of Tropical Birds

Abstract

Many audio detection algorithms have been proposed to monitor birds using their vocalizations. Among these algorithms deep learning based techniques have taken the lead in terms of performance at large scale. However, usually a lot of manual work has to be done to correctly label bird vocalizations in large datasets. One way to tackle this limitation is using the Multiple Instance Learning (MIL) framework, which models each recording as a bag of instances, i.e., a collection of audio segments that is associated with a positive label if a bird is present in the recording. In this work, we modified a previously proposed Deep MIL network to predict the presence or absence of birds in audio field recordings of one minute. We explore the behavior and performance of the network when using different number of Mel-Frequency Cepstral Coefficients (MFCC) to represent the recordings. The best configuration found achieved a 0.77 F-score over the validation dataset.