Efficient Neural Architectures for Acoustic Monitoring of Livestock

This paper evaluates the performance of three convolutional neural network (CNN) architectures, YAMNet, a VGGish-based CNN, and a custom-designed lightweight CNN, for classifying goat vocalizations. All models were assessed under the same experimental conditions, using consistent logMel spectrogram representations and a shared data augmentation strategy. The results show that all architectures achieve comparable accuracy levels above 82 %, with the VGGish-based CNN reaching the highest performance $(82.75 \%)$. However, this model also exhibits a high computational cost, requiring approximately 55 million parameters. In contrast, the custom CNN achieves a similar accuracy $(82.40 \%)$ while using only 406,408 parameters. These findings highlight the effectiveness of compact, applicationoriented CNNs, which offer strong classification performance with significantly reduced computational requirements. Such models are particularly suitable for real-time, energy-efficient deployment in resource-constrained environments typical of Precision Livestock Farming (PLF).