A tristimulus-formant model for automatic recognition of call types of laying hens

Abstract

An essential objective of Precision Livestock Farming (PLF) is to use sensors that monitor bio-responses that contain important information on the health, well-being and productivity of farmed animals. In the literature, vocalisations of animals have been shown to contain information that can enable farmers to improve their animal husbandry practices. In this study, we focus on the vocalisation bio-responses of birds and specifically develop a sound recognition technique for continuous and automatic assessment of laying hen vocalisations. This study introduces a novel feature called the “tristimulus-formant” for the recognition of call types of laying hens (i.e., vocalisation types). Tristimulus is considered to be a timbre that is equivalent to the colour attributes of vision. Tristimulus measures the mixture of harmonics in a given sound, which grouped into 3 sections according to the relative weights of the harmonics in the signal. Experiments were designed in which calls from 11 Hy-Line brown hens were recorded in a cage-free setting (4303 vocalisations were labelled from 168 h of sound recordings). Then, sound processing techniques were used to extract the features of each call type and to classify the vocalisations using the LabVIEW® software. For feature extraction, we focused on extracting the Mel frequency cepstral coefficients (MFCCs) and tristimulus-formant (TF) features. Then, two different classifiers, the backpropagation neural network (BPNN) and Gaussian mixture model (GMM), were applied to recognise different call types. Finally, comparative trials were designed to test the different recognition models. The results show that the MFCCs-12+BPNN model (12 variables) had the highest average accuracy of 94.9 ± 1.6% but had the highest model training time (3201 ± 119 ms). At the same time, the MFCCs-3+TF+BPNN model had fewer feature dimensionalities (6 variables) and required less training time (2633 ± 54 ms) than the MFCCs-12+BPNN model and could classify well without compromising accuracy (91.4 ± 1.4%). Additionally, the BPNN classifier was better than the GMM classifier in recognising laying hens’ calls. The novel model can classify chicken sounds effectively at a low computational cost, giving it considerable potential for large data analysis and online monitoring systems.