Deep learning solutions for audio event detection in a swine barn using environmental audio and weak labels

The increasing demand for animal protein products has led to the emergence of Precision Livestock Farming (PLF) and the adoption of sensing technologies, big data solutions, and Machine Learning (ML) methods in modern livestock farming. At the same time, the audio signal processing field has undergone notable advancements in recent years, transitioning from traditional techniques to more sophisticated ML approaches, with open challenges in detecting and classifying complex, low-quality, and overlapping sounds in real-world scenarios. In this paper, we evaluate deep learning methods, conceived from computer vision to attention-based approaches, for Audio Event Detection (AED) on a novel audio dataset from a swine farming environment with challenging characteristics, such as weak annotations and high amounts of noise. The primary purpose of our study is to prospect effective AED solutions for the development of tools for auditing livestock farms, which could be used to improve animal welfare. Our results show that, despite inherent limitations in the dataset’s size, class imbalance, and sound quality, Convolutional Neural Network (CNN) and attention-based architectures are respectively effective and promising for detecting complex audio events. Further research may explore avenues for optimizing model performance in similar, real-life datasets while simultaneously amplifying annotated events and reducing annotation costs, thereby enhancing the broader applicability of AED methods in diverse audio processing scenarios.