Alarming pig vocalization-based prediction using the self-supervised BEATs model

This paper presents a learning framework for classifying pig vocalizations to enhance automated animal welfare monitoring. The primary goal is to distinguish pig sounds from farm noise and to classify pig vocalizations as “alarming” or “non-alarming”. Leveraging the pre-trained BEATS (Bidirectional Encoder representation from Audio Transformers) model, we employ a two-phase training strategy that combines supervised learning with a semi-supervised approach using pseudo-labeling. The initial supervised model achieves promising accuracy. By fine-tuning the model with a large unlabeled dataset, performance is significantly enhanced. The final model for pig sound detection reaches 95.1 % accuracy, while the model for identifying alarming sounds achieves 95.8 % accuracy. A prototype application was developed and deployed on a NVIDIA Jetson Nano, demonstrating the model’s utility for real-time, on-site prediction in a barn setting. These results confirm the framework's robustness and potential for real-world application in precision livestock farming.