Learning Higher Representations from Bioacoustics: A Sequence-to-Sequence Deep Learning Approach for Bird Sound Classification

Abstract

In the past two decades, a plethora of efforts have been given to the field of automatic classification of bird sounds, which can facilitate a long-term, non-human, and low-energy consumption ubiquitous computing system for monitoring the nature reserve. Nevertheless, human hand-crafted features need numerous domain knowledge, and inevitably make the designing progress time-consuming and expensive. To this line, we propose a sequence-to-sequence deep learning approach for extracting the higher representations automatically from bird sounds without any human expert knowledge. First, we transform the birds sound audio into spectrograms. Subsequently, higher representations were learnt by an autoencoder-based encoder-decoder paradigm combined with the deep recurrent neural networks. Finally, two typical machine learning models are selected to predict the classes, i.e., support vector machines and multi-layer perceptrons. Experimental results demonstrate the effectiveness of the method proposed, which can reach an unweighted average recall (UAR) at 66.8% in recognising 86 species of birds.