BAE-LiteNet: a lightweight behavior-aware network with diffusion-based augmentation for sow estrus vocalization recognition

Accurate recognition of sow estrus vocalizations is essential for reproductive management in pig farming but remains challenged by the scarcity of samples, environmental noise, and deployment constraints. This study proposes BAE-LiteNet, which integrates diffusion-based data augmentation with a lightweight, behavior-aware recognition model. To address data imbalance, a conditional diffusion model guided by Mel spectrograms generates high-fidelity estrus audio samples, reducing annotation costs while preserving behavioral features. The recognition model employs a three-branch multi-scale convolutional front-end to capture diverse time-frequency patterns and a behavior-aware deep feedforward sequential memory network (BA-DFSMN) for estrus vocalization recognition. Spatial feature extraction is enhanced through a depthwise-pointwise convolutional module, contributing to model compactness. Evaluations on a dataset of 4,400 labeled sow vocalizations show that BAE-LiteNet achieves 98.18% accuracy and a 98.17% F1-score with only 0.75 million parameters. Comparative and ablation studies confirm the robustness of the model under noisy, imbalanced conditions. These results highlight BAE-LiteNet’s potential as a non-invasive, efficient solution for estrus monitoring in livestock farming.