Bird Call Recognition using a Convolutional Neural Network
No Thumbnail Available
Date
2021
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Uva Wellassa University of Sri Lanka
Abstract
Birds are an important indicator of biodiversity in an eco-system as well as identified as a crucial
indicator of health of an environment. They are susceptible to environmental changes therefore; it is
significant to investigate and monitor bird species. Birds have different sounds (which are called
„birds calls‟) that are melodious to the human ear. Therefore, bird call identification and classification
remains as very interesting and important area because it requires expert naturalists to manually
identify bird type according to the bird call. Furthermore, birds have different types of voices. Such
as calling, singing and mimicking representing different acoustic characteristics and this research
suggests a novel pre-tained neural network model using verbosity,number of epochs, frequency and
batch size of bird calls in identification of bird type and analysis how these factors will affect to the
accuracy of bird call prediction. This research introduces a novel and a scientific method which can
be used in bird explorations and helps ecologist to be aware of environmental changes. But there are
some challenges when doing bird call classification such as background noises, different types of bird
voices, inter-species variance and multi -label classification problem. This study use as pre-trained
Neural Network (CNN) Model with bird recordings acquired by the Xeno-canto bird sharing
database. Spectrograms are generated for every bird call and background noises were filtered. The
key fact of bird species identification is the extraction of features of bird vocalizations. And neural
network model has a strong self-learning and signal extraction ability, and it can automatically
acquire and combine characteristic information from the bird recordings. Mel-Frequency
cepstrum(MFC), Finite Impulse Response (FIR) filter and Fast Fourier Transform(FFT) techniques
were performed on each and every spectrogram and after creation of Mel-spectrogram data were split
into train process and train the dataset using Keras and Librosa libraries. The proposed methodology
predicts the birds‟ name analyzing the bird call with 90% of accuracy. This is of great significance to
the identification of birds with small sample size.
Keywords: Convolutional Neural Network; Mel-Frequency Cepstrum; Fast Fourier Transform; Finite
Impulse Response
Description
Keywords
Environment Science, Ecology, Animal Sciences, Neural Network, Biodiversity – Sri Lanka