Region-based Birdcall Recognition Using Signal Processing

The unique trait of birdsong provides a method for the investigation of bird population based on birdcall recognition. Due to the living habit and geographical distribution of bird species, the recognition of birdcall might be affected by birds' location. The effect of location on a different type of models' prediction performance is investigated in this paper. This research implements three types of predictive models - convolutional neural network (CNN), recurrent neural network (RNN), and support vector machine (SVM) - over three countries: Japan, Sweden, and Norway. Signal processing is used for pre-processing of the dataset to obtain Mel Frequency Cepstrumas, the input of three models. Based on the evaluation of model performances, including training time, accuracy, AUC score, F1 score, recall, and precision, the best model of each country is selected: SVM for Japan with 93% accuracy, CNN for Sweden and Norway with 98% and 94% accuracy respectively. The difference in the performance of models in different countries indicates that a country-based prediction of birdsong using varied models may help improve the performance of birdcall recognition.