An Auto Encoder-Decoder Approach to Classify the Bird Sounds Using Deep Learning Techniques

Nowadays identification of bird species is the main problem for ornithologists and ecologists. The increase in volumes of data labelling, still there is a scope to implement an automated bird classification through deep learning techniques. Identifying the nature of a bird and its flock aids in analyzing different other species in the natural environment. This is considered as the main role in the problem of ecology. The species of birds act accordingly to the changes in ecology and this became a challenging task to identify the growth of the bird species with advanced technologies. In this work, our aim is to classify the different bird species based on voice recordings across various locations using autoencoders and decoders. This proposed work is a four-step process: data preprocessing of bird acoustic, feature extraction using a deep network, classification by random forest (RF) and support vector machine (SVM) and identifying the pitch intensities of a bird in low, medium and high ranges. In data pre-processing, the recorded audio is converted into spectrograms through frequency waves. The Mel Spectrogram filter is used to remove the noise from the recorded bird audio. For extracting the features in the spectrogram image, encoders and decoder convolutions are used to construct a fully connected vector. These features are given to RF and SVM for classifying the bird species to obtain standard prediction accuracy.