Performance Analysis and Evaluation of LSTM and GRU Architectures for Houston toad and Crawfish frog Call Detection

Audio signal analysis has become prominent in biological domains toward applications in detecting endangered or threatened species like Houston toad and Crawfish frog. Researchers at Texas State University and Texas A&M University are working on a project to rescue these species and understanding the causes of their decline. Currently the researchers are using an Automated Recording Device (ARD), Toadphone 1, an embedded solution designed for only Houston toad call detection. However, this device's software solution has shown limited success in identifying toad calls consequent of high false-positive rates. This paper experimented with a modified software solution for existing ARD, which is capable of detecting Houston toad and Crawfish frog calls with decreased false-positive rates. Six experiments to detect the calls were designed by using thirty-nine Mel-Frequency Cepstral Coefficients (MFCCs) with delta and delta-delta coefficients and sixteen Spectral Sub-band Centroids (SSCs) as audio features within Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRUs) as the classifiers. Results show that LSTM as the classifier with thirty-nine MFCCs audio features, and a 20% validation split produces the highest accuracy for detecting Houston toad and Crawfish frog calls. This architecture has gained 84.7% training, 82.05% validation accuracy, and 84.2% test accuracy with 91.4% test accuracy on Houston toad call and 77.1% on Crawfish frog call.