Designing a Lightweight Convolutional Neural Network for Bird Audio Detection

State-of-the-art machine learning models have been used for a plethora of Sound Event Detection (SED) applications, including bird vocalization detection, which plays a significant role in monitoring the overall ecosystem health. However, many popular models are characterized by high complexity and, therefore, are rather unsuitable for operating on IoT edge devices. To address this limitation, this paper proposes a lightweight Convolutional Neural Network with emphasis on edge computing. The proposed architecture achieves an accuracy measure of 86.42% on a bird audio detection task with only 73,377 trainable parameters.