The goal is to classify different birds species based on their songs/calls. Spectrograms have been extracted from the audio samples and used as features for classification. A Siamese network based on 1D dilated convolutions is used here. Model is trained using triplet loss. I am using the British Birdsong Dataset available on Kaggle for this experiment.
Download the data from here. This dataset is a subset of Xeno-Canto database. Siamese Networks along with dilated 1D convolutions are used here to classify 9 different bird species.
Confusion Matrix of testset:
Scatter plot of embeddings after applying PCA:
Note: If you are having this error: AttributeError: module 'keras.utils.generic_utils' has no attribute 'populate_dict_with_module_objects'.
Type:
pip uninstall tf-nightly
pip uninstall tf-estimate-nightly
pip install tensorflow --upgrade --force-reinstall