Artificial Neural Network approach to assess vocal identity, kinship and ageing in goats (Capra hircus)

Machine learning techniques are becoming a more important tool in studying animal vocal communication. The goat (Capra hircus), is a very social species where vocal communication and recognition, particularly between mothers and offspring, is important. In this study we tested the reliability of a Multi-Layer Perceptron (feed-forward ANN) to automate the process of vocal analysis for individuality, group membership and ageing in this species. Vocalisations were obtained from 10 half-sibling (same father but different mothers) goat kids, belonging to 3 distinct social groups. We recorded 164 contact calls emitted during early postnatal days, and 157 additional calls recorded from the same individuals at 5 weeks. For each call we measured 27 spectral and temporal acoustic parameters using automatized procedures in PRAAT (www.fon.hum.uva.nl/praat). For each classification task we built stratified 10-fold cross-validated neural networks using the WEKA software package (www.cs.waikato.ac.nz/ml/weka). The input nodes corresponded to the acoustic parameters measured on each signal. ANNs were trained with the error-back-propagation algorithm. The number of hidden units was set to the number of attributes + classes. Each model was trained for 300 epochs (learning rate 0.2; momentum 0.2). To estimate a reliable error of the models, we repeated 10-fold cross-validation iterations 10 times and calculated the average predictive performance. The Correctly Classified Instances was 71.13±1.16% for the vocal individuality, 79.59±0.75% for the social group and 91.37±0.76% for the age of the vocaliser. Our results demonstrate that ANNs are a powerful tool for studying contact calls. The performances we achieved are higher than those previously obtained with classical statistical methods such as Discriminant Function Analysis. Further development of this approach might include the classification of contact calls within other social species and the comparison of ANNs with other machine learning techniques.