Deep neural network approach to frog species recognition

Automatic frog species recognition based on acoustic signal has received attention among biologists for environmental studies as it can detect, localize and document the declining population of frog species efficiently compared to the manual survey. In this study, we investigate the possibility of the use of Deep Neural Network (DNN) as a classifier for a frog species recognition system. The Mel-Frequency Cepstral Coefficients (MFCCs) is utilized as features and prior to the feature extraction, we also investigate the capability of automatic segmentation of syllables based on the Sinusoidal Modulation (SM), Energy with Zero Crossing Rate (E+ZCR) and Short-Time Energy with Time Average Zero Crossing Rate (STE+STAZCR). We also evaluate several DNN parameter's setting so as to discover the optimum parameter values for our developed system. 55 different species of frog with 2674 syllables from our in-house database have been tested. Experimental results based on DNN classifier showed that the STE+STAZCR method gives the accuracy of 99.03%, which reveals the viability of DNN as a classifier. In future, further research on DNN parameter optimization will be conducted for system improvement.