A classifier-deduced signal extraction approach for time difference estimation in acoustic sensor networks

With the development of reliable AI-based species classifiers and the design of low-cost autonomous recording units, acoustic monitoring has become an emerging research field. Although strides are made in automated species monitoring, automated localization remains a significant challenge. Distinguishing and pinpointing bird sounds in noisy, reverberant, and dynamic natural environments is extremely difficult, ultimately deteriorating the accuracy of time difference estimations and, consequently, localization. In this paper, we take a significant step towards reliable automated localization by presenting a viable and generalizable approach to extracting species-dependent signals from intermixed acoustics, which we call Classifier-Deduced Signal Extraction (CDSE). These signals can be used to estimate precise time differences while retaining information for individual species. Our method seamlessly extends the current capabilities, requiring only minor modifications to state-of-the-art classifiers. We prove its applicability and usefulness by deploying it on bird acoustics using the popular bird species classifier BirdNET. • Novel signal extraction method for species sounds in adverse acoustic environments. • Demonstration how today’s AI-based species classifiers for birds can be modified. • Proof-of-Concept evaluation on bird acoustics including time difference estimation.