Acoustic Data Augmentation for Small Passive Acoustic Monitoring Datasets

Abstract

Training complex deep neural networks can result in overfitting when the networks are trained from random weight initialization on small datasets. Data augmentation helps to reduce the negative effects of overfitting. Data augmentation is the process by which the amount of data for a given problem is increased in quantity via some augmentation technique. The findings in computer vision and audio recognition research reveals that the performance of machine learning classifiers is significantly improved when the data is augmented. In the context of ecology, researchers conduct field surveys whereby microphones are placed in some location and audio data is recorded over a period of time. There is however no guarantee that the particular species of interest in the field survey will vocalize frequently near the microphone. Thus, the amount of data captured for the species of interest might be limited. Training robust classifier models on such limited data will most likely lead to overfitting. The purpose of this research is to investigate several audio augmentation techniques as a means to increase the amount of audio examples for certain species of interest with the goal of creating robust audio vocalization classifier models. We investigate noise injection and time and frequency masking data augmentation techniques. These techniques are applied to two birds of interest, namely the pin-tailed whydah (Vidua macroura) and the Cape robin-chat (Cossypha caffra). While these two species are not endangered, they allow us to compare the various augmentation techniques. The audio recordings were obtained from the Intaka Island Nature Reserve, South Africa. To evaluate the performance of the augmentation techniques we conducted a com parison between experiments run with and without augmentation. We chose to use convolutional neural networks as our classifier given that they are the state-of-the-art in audio recognition tasks. Furthermore, convolutional neural networks have revealed good performance in the field of bioacoustics. We manually annotated 768 audio files (20 minutes each) totaling over 256 hours of audio