Animal vocalizations can differ depending on the context in which they are produced and serve as an instant indicator of an animal's emotional state. Interestingly, from an evolutional perspective, it should be possible to directly compare different species using the same set of acoustic markers. This paper proposes a deep neural network architecture for analysing and recognizing vocalizations representing positive and negative emotional states. Understanding these vocalizations is critical for advancing animal health and welfare, a subject of growing importance due to its ethical, environmental, economic, and public health implications. To this end, a framework assessing the relationships between vocalizations was developed. Towards keeping all potentially relevant audio content, the constructed framework operates on log-Mel spectrograms. Similarities/dissimilarities are learned by a suitably designed Siamese Neural Network composed of convolutional layers. The formed latent space is appropriately clustered to identify the support set facilitating the emotion classification task. We employed a publicly available dataset and followed a thorough experimental protocol. The efficacy of such a scheme is shown after extensive experiments considering both classification and support set selection. Last but not least, by elaborating collectively the network's activations when processing positive and negative vocalizations, important differences in the time-frequency plane are evidenced across emotions and species, assisting their understanding from animal scientists.
Stavros Ntalampiras (Wed,) studied this question.