Classification of a small mammalian vocalization data set using deep learning techniques

Activity: Talk or presentationTalk or presentation at a conference


In recent years, the automatic extraction of bioacoustic from data sets has been facilitated by artificial intelligence methods. Different machine learning frameworks have allowed to automate feature extraction and classification, which, in turn, has led to improved call identification and species recognition. However, whether the application of these types of algorithms is successful, is partly determined by the amount of available training data. In the field of bioacoustics, researchers often deal with small datasets of vocalizations. This is a problem for machine learning as a small dataset can lead to problems such as overfitting and lack of generalization (i.e., poor performance on out-of-sample data). Here, we show how to tackle the problem of small datasets in a bioacoustic classification task using neural networks. We present and explain how to use techniques such as pre-training and data augmentation. We specifically emphasize how to use them in a meaningful way so as to not alter the nature and distribution of the features (for instance, fundamental frequency) that are relevant to the classification task. To illustrate the effect of these techniques, we present an example of a classification of calls from a database of phylogenetically distant mammal calls that is limited in the number of calls it contains. For this purpose, we use a well-known pre-trained convolutional neural network that has been successfully applied in prior bioacoustics tasks. We can show that with the correct techniques, sufficient accuracy can be obtained across species. To further extend the analysis, we also trained the neural networks using specific combinations of species for training and testing. With this approach, we can explore how features from different species vocalizations affect the accuracy across two affiliative conditions. Finally, we discuss the possibility that these results may be similar to human perception, comparing human and machine classification.
Period28 Oct 2023
Event titleThe XXVⅢ International Bioacoustics Congress
Event typeConference
LocationHokkaido, Japan
Degree of RecognitionInternational