PigBehaviorVocal Dataset (PBV-5)

All animal data collection procedures in this study complied with relevant guidelines for laboratory animal welfare and ethical review and were approved by the Animal Ethics Committee of Henan Institute of Science and Technology (Approval No. LLSC2025107). Based on this, a pig vocalization dataset was constructed from real commercial farm environments for multi-class behavioral sound recognition. The dataset comprises 5,268 samples covering five typical behavior categories, including cough (1,053 samples), scream (1,092 samples), estrus (1,076 samples), feeding (1,029 samples), and normal sounds (1,018 samples), with an average duration ranging from 1.75 to 1.94 s across categories.

During data preparation, continuous recordings were first segmented into 5–10 s clips for event localization and preliminary annotation, followed by further trimming into 1–2 s event-centered segments to form the final samples. All samples were manually annotated and subsequently reviewed; initial annotations were performed by experienced farm personnel based on husbandry records and on-site observations, and then verified by licensed veterinarians to improve labeling accuracy and consistency. After preprocessing, annotation, and feature extraction, all samples were further organized into structured JSON files. Each JSON file corresponds to an individual sample and stores the behavior label, the final quantized feature representation, and necessary auxiliary information, enabling efficient data loading, model training, and evaluation.

This dataset is intended solely for non-commercial use in research and educational contexts. For further details regarding data field definitions, JSON file organization, sample partitioning, or additional supporting materials, please contact the corresponding author, Jianping Wang (wangjianping@hist.edu.cn).