Sound-based animal recognition model

Abstract

Conservation of biological and biological diversity is the goal of joint efforts worldwide, and protecting wild creatures requires various monitoring equipment with identification capabilities. In this case, this article will introduce a model with animal call recognition ability designed using edge impulse. A two-layers convolutional neural network (CNN) is selected and trained with the input features generated by Mel Frequency Cepstral Coefficient, Spectrogram and Mel-Spectrogram. The model hoped not only to focus on the recognition accuracy rate of the model but also on the inferencing time, peak RAM usage and other performance to cope with the poor performance of the tiny equipment used in the real wild. Using five classes consisting of over 4,000 audio samples as the input and after several round experiments. Finally, the MFCC&CNN model is chosen because it has an accuracy rate of over 85.6%, the swiftest inferencing time and the lowest peak RAM usage.