CNN-based Multi-model Birdcall Identification on Embedded Devices

Shidong Pan, Dehai Zhao, Weishan Zhang

Abstract: Bird species identification by their vocalization on embedded devices is an important and challenging task. In this paper, we propose BIRD (Bird Identity Record and Detection) system as a complete IoT solution for birdcall identification in wild field environment, and verify its feasibility by a simulation experiment. The bird call signals are firstly converted to spectral field then input into deep neural network to classify bird species. To improve the classification performance in BIRD system, we introduce a CNN (Convolutional Neural Network)-based multi-model network, which fuses acoustic signals and geographical coordinate information into a unified model. Our work achieves 0.849 accuracy and 0.749 F 1-score over 397 species on BirdCLEF2021 dataset, outperforming traditional classification models.