Selective audio recording device for wildlife research using embedded machine learning

Abstract

Wildlife monitoring through sound recording has become an essential tool in ecological research. However, challenges such as limited power and memory constraints hinder large-scale, long-term deployment of monitoring devices. To address these limitations, this paper presents a novel wildlife monitoring device that integrates embedded machine learning (ML) for event-triggered recording. This system captures only relevant sounds, leading to a more efficient memory usage and power consumption than the traditional fixed-schedule scheme, and a significantly larger percentage of useful data collected. The device features a low-cost, low-power hardware design equipped with a digital microphone, dual MicroSD storage and a flexible power system. Its embedded ML component enables real-time audio classification and selective recording triggered by specific acoustic events. Preliminary testing using a prototype device demonstrated effective detection of penguin vocalizations, achieving an average current intake ranging from 4.06 to 6.02 mA, depending on the operational mode. This enables the device to be powered by a small, cost-effective rechargeable battery and solar power, supporting near-perpetual operation. The proposed system represents a step forward in deploying low-cost, low-power scalable devices for acoustic wildlife monitoring.