Individual Animal Identification through Bioacoustics
The possibility of accurately identifying single animals based on passive recording of their vocalizations can revolutionize our understanding of animal behavior within populations and communities. Moreover, understanding how individuals within a population communicate can provide valuable knowledge to improve conservation models. However, current passive animal acoustic monitoring literature does not offer identification of specific individuals, obscuring the full potential of studying authentic animal interactions and behaviors. Recent advances in the field have demonstrated the potential of deep-learning (DL) methods to identify multiple species from bioacoustic data. Typically, this process involves using well-established Convolutional Neural Network (CNN) architectures with spectrograms as input. While these studies have paved the way for the application of DL in bioacoustics, there remains a lack of comprehensive validation across diverse datasets. In this study, we propose extending these efforts by developing a deep-learning model that can accurately identify vocalizations of individual animals within the same species. Our approach utilizes audio recordings from various databases, including those recorded in controlled conditions, and incorporates synthetic data to enhance model training. Specifically, we aim to leverage the benefits of synthetic data, which includes labeled vocalizations from individual animals to pre-train deep neural networks. These pre-trained networks will subsequently be fine-tuned using real-world data, with the expectation that this approach will lead to improved model performance.