MFCC-based Houston Toad Call Detection using LSTM

Abstract

The Houston Toad (HT) is an endangered amphibian living in the edge of extinction. For their conservation, the localization of their mating call needs to be detected in order to protect the eggs from being hunted by predators. Due to the remote locations of their habitat, solar-powered Automatic Recognizing Device (ARD) has been deployed to identify the distinct HT mating call at Bastrop, Texas. The ARD records environmental sound at prescribed intervals that implements signal processing and trained Multilayer Perceptron (MLP) Neural Network (NN) predictor model to identify the HT onboard. If an HT exists, the ARD sends notification of timestamp via Email and SMS over the GPRS modules to the researcher. The signal processing techniques applied to the audio file are band-pass filtering, framing, windowing, clipping, feature extraction by Mel-Filterbank, and Mel-Frequency Spectral Coefficient (MFCC). The MLP-NN resulted in 66.67% success rate on detecting HT calls. This paper modifies and tunes the signal process techniques by modifying the band-pass filter, frame size, only using MFCC and implements Long Short-Term Memory (LSTM) classifier that results in success rate of 98% on true-positive, and 86% on true-negative over 94% training and 92.6% testing accuracy for HT detection.