Classification of bird species based on vocalization audio recordings using Convolutional Neural Networks
The problem of bird vocalization classification is well known and has been addressed numerous times. Specifically in Data Science papers, it has appeared repeatedly over the years with progressively better attempts. One of succesful approaches is using Convolutional Neural Network with visual representation for audio signal (e.g. (mel) spectrograms) as features. Such approach is also implemented in this project.
Main project file (Jupyter Notebook): birdify.ipynb
Library containing all core preprocessing methods can be found under: /soundmatrix
Csv file containing information about samples: dataset.csv
Python script for resampling and signal/noise splitting: sample_gen.py
pandas
numpy
librosa
pydub
scipy
tensorflow
tensorflow-io
All audio data used in this project come from Bird songs from Europe (xeno-canto) collection built from a xeno-canto API query and uploaded on Kaggle by Francisco de Abreu e Lima.
[1] Sprengel, E., Jaggi, M., Kilcher, Y., & Hofmann, T. (2016). Audio Based Bird Species Identification using Deep Learning Techniques. CLEF.
[2] Kahl, S., Wilhelm-Stein, T., Hussein, H., Klinck, H., Kowerko, D., Ritter, M., & Eibl, M. (2017). Large-Scale Bird Sound Classification using Convolutional Neural Networks. CLEF.
[3] Smith, L. N., (2017). Cyclical Learning Rates for Training Neural Networks. IEEE Winter Conference on Applications of Computer Vision.
[4] Lasseck, M., (2013). Bird song classification in field recordings: winning solution for nips4b2013 competition. Proc. of int. symp. Neural Information Scaled for Bioacoustics.
[5] Abreu e Lima, F., (2020). Bird songs from Europe (xeno-canto). https://www.kaggle.com/dsv/1029985