Pig cough detection using deep features and an improved SKA-TDNN model.

Pig coughing is an important acoustic indicator for the early detection of respiratory diseases in swine. Traditional monitoring relies heavily on manual inspection, which is labour-intensive and increases the risk of cross-infection. Therefore, intelligent detection of cough sounds using audio-based methods is essential for improving disease prevention and breeding efficiency. However, most existing studies focus on traditional audio features, such as Mel-Frequency Cepstral Coefficients (MFCCs) and filter bank (F-bank) features, which often struggle to maintain recognition accuracy in the complex acoustic environments of pig farms. This study proposes a novel framework that employs deep feature representations as an alternative to handcrafted features, thereby capturing more robust acoustic patterns. In addition, multiple data augmentation techniques are applied to enhance data diversity and model generalisation. Building on the Time Delay Neural Network (TDNN) architecture, we further design a Simplified Kernelized Attention TDNN (SKA-TDNN) model, which integrates a lightweight attention mechanism to improve temporal feature modelling while significantly reducing the number of parameters. Experimental results show that models trained on deep features outperform those based on MFCC and F-bank features under various evaluation metrics. When compared against mainstream architectures including Convolutional Neural Networks (CNNs), ECAPA-TDNN, and conventional TDNNs, the proposed SKA-TDNN achieves the best performance, reaching an overall accuracy of 98.9%, with only 5.83 MB of parameters. These findings highlight the novelty and practical value of introducing deep features with a lightweight attention-enhanced TDNN for animal cough detection. Beyond swine, the proposed framework provides a promising and generalisable approach for intelligent respiratory disease monitoring in other livestock and domestic animals. Moreover, the system is particularly suited for deployment in modern large-scale farming environments, where automated, real-time health monitoring is essential for precision livestock management.