Sound analysis and detection, and the potential for precision livestock farming - a sheep vocalization case study

Abstract

Livestock vocalizations contain a wealth of information pertaining to welfare state and behaviour. Acoustic monitoring is non-invasive and has potential for numerous Precision Livestock Farming (PLF) applications. A key step in the development of a PLF acoustic monitoring system is the development of stock vocalization detection and classification algorithms. To this end, an algorithm based on Mel-Frequency Cepstral Coefficients (MFCCs) and Support Vector Machines (SVMs) was created. Audio data was acquired from a sheep farming enterprise, reflecting realistic operating conditions. Algorithm performance was across three experiments: (i) sheep vocalization classification, (ii) adult vs. juvenile classification, (iii) multi-animal vocalization. Performance in experiments (i) and (ii) was very high (>98% accuracy, stratified 10-fold cross-validation). A novel probability-based approach is proposed to handle the difficult problem of experiment (iii). The use of a threshold allows application-specific customization of class classification distribution. By use of the MFCC-SVM algorithm it is entirely possible to detect and classify sheep vocalizations in noisy environments. These results, combined with examples from the literature, show that sound analysis and detection holds promise for PLF.