Development of a TinyML-based Device for Automatic Detection of Poultry Diseases using Chicken Vocalizations

Abstract

Poultry farming provides essential protein but faces challenges from diseases that affect bird welfare and farm economics. Traditional detection methods are slow and costly, underscoring the need for real-time, low-cost solutions enabled by precision livestock farming technologies. This paper presents a TinyML-based device for the automatic detection of poultry diseases by analyzing chicken vocalizations. The proposed system uses edge computing on a microcontroller to process audio data on-site, without requiring a constant internet connection or powerful hardware. Key steps include noise reduction with FIR filters, endpoint detection via a double-threshold short-time energy method, and feature extraction using Mel-Frequency Energy. A lightweight neural network with 1D-CNN and DNN layers classifies sounds as healthy, unhealthy, or noisy. We developed a prototype on the ESP32-S3 MCU, achieving 90.58% test accuracy and 0.98 F1-score for unhealthy detection. The device processes 3-second audio in 380 ms of preprocessing and 97 ms per inference window, using only 419 KB of flash and 258 KB of RAM. This low-cost solution enables scalable IoT integration for early disease intervention in poultry farms.