Robustness of CNN-Based model assessment for pig vocalization classification across diverse acoustic environments

• Robust pig vocalization classification model using Mixed-MMCT features and CNN. • Built datasets from three commercial pig farms for robustness evaluation. • Proposed a uniform 3-seconds segmentation method for real-time sound processing. • Achieved up to 91% average precision, recall, and F1-score in robustness evaluation. The study on pig vocalization classification has attracted the attention of many researchers. However, since it arises from different environmental conditions of pig farms, developing a robust model that works well across various farms remains a significant challenge. To address these challenges, this study employs a convolutional neural network (CNN) architecture to assess the effectiveness of a pig vocalization classification model in diverse farm environments. For the experiment, three datasets named GJDS, ISDS, and JEDS were collected from three actual pig farms. Each dataset consists of 4,000 audio files (1,000 non-vocal, 1,000 coughing, 1,000 screaming, and 1,000 normal), each with a duration of three seconds. A CNN with three convolutional layer blocks, max-pooling, and two fully connected layers is trained using spectrogram features. This study employed a 10-fold cross-validation to evaluate the performance of the model. Through rigorous experiments, the model achieved accuracy of 97.68%, 99.18%, and 97.45% on the GJDS, ISDS, and JEDS datasets, respectively. Robust experiments were conducted to evaluate the effectiveness of the model on an unseen dataset by using the combination of two datasets as a training set and one as a testing set. The average performances of robust experiments in terms of precision, recall, F1-score, and accuracy were 91.96%, 91.36%, 91.66%, and 91.32%, respectively. The model shows high reliability in classifying pig vocalizations within farm environments. Although performance drops under unfamiliar acoustic conditions, these findings point the way toward future improvements in model generalization.