A Houston Toad Call Detection Initial Approach Using Gated Recurrent Units for Conservational Efforts

Abstract

Conservation management of endangered amphibians requires efficient and consistent detection. Consequently, detection of species using automatic animal voice detection from audio recordings is a topic of interest in bioacoustics. This is necessary for amphibian population stewardship as well as assessing the health of those natural systems. The Houston Toad is an endangered chorusing amphibian species, and researchers of the Biology Department at Texas State University are working on a project to prevent its extinction. The researchers' initial approach is an Automated Recording Device (ARD), Toadphone-1, which is an embedded solution. It has shown limited success in identifying toad calls. If a species is not Houston Toad but has a frequency spectrum close to Houston Toad, then Toadphone-1 falsely identifies it as a Houston Toad. Hence, the current ARD solution produces high false-positives. This paper experimented with a modified software solution for existing ARD using 39 Mel-Frequency Cepstral Coefficients (MFCCs) with delta and delta-delta coefficients as audio features and Gated Recurrent Units (GRUs) as a classifier to detect Houston Toad. Results show that this experimented software solution produces 98.82% training accuracy and 97.50% validation accuracy. Test accuracy for detecting Houston Toad is 88.57%, which is approximately 20% greater than the accuracy presented by the existing software solution of Toadphone-1.