Birdsong Classifier

General information

The library it's part of a larger research project called PiaffNet. Our goal is to make it easier for everyone to study and identify birds by their vocalizations using modern Deep Learning algorithms. PiaffNet uses modern standard of implementations in Python language. We priveledge simplicity and good enough accuracy over unnescessory complexity.

Introduction

The birdsong classifier library provides a simple predictive model based on supervised learning over data samples of annotated bird sounds. We use a Convolutional Neural Network (CNN) models. Note that this version implemented in TensorFlow library.

The classifier uses a MEL sounds spectograms with $64$ bands and a break frequency at $1750$ Hz. The spectrogam constructed using a Fast Fourier Transform (FFT) with $512$ samples at $48$ kHz sampling rate and an overlap of $25$%. The frequency range of spectogram is limited between $150$ Hz and $15$ kHz to assure the frequency range covering of the majority of bird vocalizations.

install

pip install

pip install --upgrade pip
pip install -r requirements.txt
pip list

how to use

Train your model

Preparing audio for preprocessing

When in project directory

mkdir raw_data
mkdir raw_data/train_audio
mkdir raw_data/split_data
mkdir raw_data/images_png

Copy your files in the train_audio directory, the program uses audio_dataset_from_directory so it needs to be in this format :

train_audio/
 species1/
  song1.audio
  song2.audio
  ...
 species2/
  song1.audio
  song2.audio
  ...
 ...

Slice your audio and detect silent segments :

make run_slicing

Transform audio to mel spectrogram

make run_preprocess

Train the model on your data

make run_train

Predict a bird from an audio file

make run_predict