An Empirical Evaluation of Machine Learning Approaches for Species Identification through Bioacoustics

In this paper, we investigate a complete system for identifying species from audio files. Audio data from both high quality and low quality sound files with varying degrees of background noise are collected and preprocessed for enhancing the learning capability of machine learning models. Then, fine-scale features are extracted to quantify acoustic properties of audio streams. Based on the proposed system, we evaluate a set of popular machine learning approaches on audio data from the cat and dog families. The experimental results show that the use of appropriate quality data and machine learning models yield compelling identification accuracy of species with limited user assistance.