Automated detection of broiler vocalizations a machine learning approach for broiler chicken vocalization monitoring

Abstract

The poultry industry relies on highly efficient production systems. For sustainable food production, where maintaining broiler welfare is crucial, it is essential to have robust data collection systems and automated methods for assessing broiler health and welfare. This paper presents the development and implementation of an acoustic system designed to detect and differentiate between four distinct vocalizations of broiler chickens—pleasure notes, distress calls, short peeps, and warbles—while filtering out background noise and other vocalizations. The vocalization detector is designed as a convolutional neural network with 11 two-dimensional convolutional layers and one one-dimensional convolutional layer. For training, a manually labeled vocalization library was built (>2k samples, with a total duration of 190 minutes), based on a large set of continuous audio recordings of ten male Ross 308 broiler chicks aged from 1 to 36 days. An extension with a subset of the AudioSet dataset was made to include background sounds. With this library, an overall balanced accuracy of 91.1 % was achieved by the neural network-based recognizer.