Generative AI-based data augmentation for improved bioacoustic classification in noisy environments

Overview

This repository implements generative AI-based data augmentation techniques for improving bioacoustic classification in noisy environments, particularly for bird species detection at wind farm sites. We explore Auxiliary Classifier Generative Adversarial Networks (ACGAN) and Denoising Diffusion Probabilistic Models (DDPMs) to synthesize spectrograms, enhancing training data diversity without requiring extensive expert labeling.

Dataset

The project includes a new audio dataset of 640 hours of bird calls recorded at wind farm sites in Ireland. Approximately 800 samples are expert-labeled, providing a challenging benchmark due to background wind and turbine noise.

Installation

1. Clone the repository:

   git clone https://github.com/gibbona1/SpectrogramGenAI.git
   cd SpectrogramGenAI

2. Install dependencies:

   pip install -r requirements.txt

   Ensure you have Python 3.8+.

Usage

Data Augmentation with Generative Models

1. Prepare Data: Either .wav files or 256x256 spectrograms

2. Train ACGAN:

   python train_acgan.py

3. Train DDPM:

   python train_ddpm.py

4. Generate Synthetic Spectrograms:

   python generate_spectrograms.py --model ddpm --num_samples 1000 --output_path data/synthetic/

Quality Metrics

Inception Score

python inception_score.py <image_folder>

Fréchet Inception Distance

pip install pytorch-fid
python -m pytorch_fid folder1 folder2

Fréchet Audio Distance

python frechet_audio_distance.py --bg_dir folder1 --eval_dir folder2

Classification Training

Train all classifiers on combined real and synthetic data:

python train_classifiers.py

For more details, refer to the arXiv preprint.