Continuous monitoring of livestock vocalizations offers a non-invasive tool for welfare assessment, but deploying current deep learning models in resource-constrained farm environments remains challenging due to high computational demands. This study proposes a feature-based machine learning pipeline optimized for edge computing to classify caprine vocalizations. Using the VOCAPRA dataset, which comprises 4147 labeled caprine vocalizations categorized into eight distinct welfare states and contexts, a hybrid feature extraction framework was applied to derive 156 spectral, temporal, and bioacoustic descriptors. Dimensionality reduction and a comprehensive comparative screening of 18 algorithms identified the CatBoost Classifier and a Multilayer Perceptron (MLP) as the optimal models. The CatBoost ensemble achieved a robust accuracy of 85.2%, while the optimized MLP reached 87.2% overall accuracy. An edge deployment benchmark revealed that the MLP was the best candidate with for real-time application, featuring a memory footprint of just 0.639 MB and near-instantaneous inference speeds of under 0.005 milliseconds per sample. Furthermore, feature importance and SHAP analyses revealed that mel-frequency cepstral coefficients heavily drove model decisions, particularly for identifying extreme physical distress and maternal reunion. The proposed methodology achieves competitive classification performance while dramatically reducing pre-processing and computational loads compared to image-based deep learning approaches, demonstrating the viability of lightweight-model, energy-efficient, real-time bioacoustic monitoring for precision livestock farming.
Méndez et al. (Sat,) studied this question.