Acoustic-based models to assess herd-level calves' emotional state: A machine learning approach

Abstract

Animal bioacoustics is an important tool for monitoring different aspects of the physiology, behavior and well-being of animals remotely, non-invasively and continuously. Studies using this science are growing mainly due to the development of machine learning. This work aims to investigate the use of machine-learning classifiers to determine whether calves’ vocalization audio data can be used to assess their welfare condition regarding feeding. Firstly, we collected several calves’ vocalization audio data before and after feeding the animals at different day times and ages. Then, eleven time-domain, frequency-domain and sound quality-based metrics were extracted from these audio data and used as features for the classifiers. These features were used to determine whether vocalization audio data belonged to before or after feeding classes. Moreover, the most relevant ones were identified using the Random Forest algorithm. Finally, seven machine-learning classifiers were trained and tested, considering the entire set of features and a subset containing the most relevant features. The k-nearest neighbor classifier trained with the subset of the most relevant features obtained a 98.37% accuracy. Both frequency-domain and sound quality features played important roles in this classification. The main implications of this study are the development of a methodological proposal to study acoustics using machine learning and the fact that vocalization is a biomarker of animal welfare.

Keywords

• Calves’ vocalization
• Machine-learning classifiers
• Frequency-domain-based metrics
• Sound quality domain metrics