Bioacoustic classification of a small dataset of mammalian vocalisations using deep learning

In the last few years, automatic extraction and classification of animal vocalisations has been facilitated by machine learning (ML) and deep learning (DL) methods. Different frameworks allowed researchers to automatically extract features and perform classification tasks, aiding in call identification and species recognition. However, the success of these applications relies on the amount of available data to train these algorithms. The lack of sufficient data can also lead to overfitting and affect generalisation (i.e. poor performance on out-of-sample data). Further, acquiring large data sets is costly and annotating them is time consuming. Thus, how small can a dataset be to still provide useful information by means of ML or DL? Here, we show how convolutional neural network architectures can handle small datasets in a bioacoustic classification task of affective mammalian vocalisations. We explain how these techniques can be used (e.g. pre-training and data augmentation), and emphasise how to implement them in concordance with features of bioacoustic signals. We further discuss whether these networks can generalise the affective quality of vocalisations across different taxa.