Classification of Goat Vocalization via Lightweight Machine Learning and High-Dimensional Acoustic Features

Area of Science:

• Animal Science
• Bioacoustics
• Machine Learning

Background:

• Livestock vocalizations offer non-invasive welfare assessment, but deep learning models face computational challenges in farm settings.
• Resource-constrained environments require efficient, low-power monitoring solutions for precision livestock farming.

Purpose of the Study:

• To develop and evaluate a feature-based machine learning pipeline for classifying caprine vocalizations suitable for edge computing.
• To identify optimal algorithms and features for real-time, low-resource bioacoustic monitoring of goat welfare.

Main Methods:

• Utilized the VOCAPRA dataset with 4147 labeled caprine vocalizations across eight welfare states.
• Applied a hybrid feature extraction framework generating 156 spectral, temporal, and bioacoustic descriptors.
• Screened 18 algorithms, including CatBoost and Multilayer Perceptron (MLP), for classification and edge deployment.

Main Results:

• The optimized MLP achieved 87.2% overall accuracy, outperforming the CatBoost ensemble (85.2%).
• The MLP demonstrated excellent edge-computing viability with a 0.639 MB memory footprint and <0.005 ms inference speed.
• Mel-frequency cepstral coefficients were key features, particularly for identifying distress and maternal reunion.

Conclusions:

• A feature-based machine learning approach offers a viable, computationally efficient alternative to deep learning for livestock bioacoustic monitoring.
• The proposed lightweight MLP model is suitable for real-time, energy-efficient welfare assessment in precision livestock farming.
• This methodology significantly reduces computational load, making advanced monitoring accessible in resource-limited farm environments.