Speaker
Description
Effective management of respiratory diseases in cattle (Bos taurus), particularly within the context of Bovine Respiratory Disease Complex (BRDC), relies on early, non-invasive monitoring strategies to safeguard animal health and welfare. Among the earliest observable indicators, vocal and respiratory-related sounds such as coughs and vocalizations provide valuable insight into physiological and behavioural states. These acoustic cues offer strong potential for continuous, automated surveillance in precision livestock farming systems. In this study, we present a deep learning-based framework for the detection of cow vocal activity using environmental audio recordings collected from multiple farm settings. The proposed system employs convolutional neural networks operating in real time to identify vocal events. Unseen recordings were used to evaluate the robustness and generalization capabilities of the approach.Results indicate that incorporating data from heterogeneous farm environments enhances model generalization, achieving an F1-score of 57.40% and a recall of 74.05%. While models trained on single-farm data can reach higher peak performance under matched conditions, cross-farm training yields more stable behaviour across varying acoustic contexts. Additionally, detection performance is shown to depend strongly on temporal segmentation strategies, highlighting the importance of parameter selection for deployment. This work demonstrates the feasibility of scalable, non-invasive acoustic monitoring systems for livestock, supporting early detection of health and behavioural changes. Such approaches contribute to the development of autonomous tools for improving animal welfare and farm management practices.