OrcaCallGAN: Spectrogram-Based GANs for Orca Call Synthesis

A SpecGAN-based deep learning framework for generating realistic orca vocalization spectrograms from bioacoustic recordings.

---

Project Overview

This project explores the application of Generative Adversarial Networks (GANs) to bioacoustic signal synthesis. The model learns temporal and spectral characteristics of real orca vocalizations and generates realistic spectrograms from latent vectors.

The complete pipeline includes:

• Audio preprocessing
• Noise reduction
• Spectrogram generation
• Data normalization
• Data augmentation
• SpecGAN training
• Spectrogram synthesis
• Audio reconstruction

---

Sample Outputs

---

Performance Metrics

Metric	Value
Dataset Size	732 audio samples
Architecture	SpecGAN
Framework	PyTorch
FID Score	46.3
Device	Apple Silicon MPS
Audio Domain	Orca Vocalizations

---

Pipeline

Raw Audio Recordings
          ↓
Noise Reduction
          ↓
Mel Spectrogram Generation
          ↓
Normalization
          ↓
Data Augmentation
          ↓
SpecGAN Training
          ↓
Generated Spectrograms
          ↓
Audio Reconstruction

---

Dataset

The dataset consists of 732 orca vocalization recordings collected from publicly available marine bioacoustic datasets.

Preprocessing Steps

• Noise reduction
• Equalization filtering
• Mel spectrogram generation
• Normalization
• Data augmentation

---

Model Architecture

Latent Noise Vector
          ↓
      Generator
          ↓
Generated Spectrogram
          ↓
    Discriminator
          ↓
Real / Fake Classification

---

Repository Structure

├── Preprocessing/
├── Results/
├── generated_audio/
├── dataset.py
├── model.py
├── losses.py
├── train.py
├── invert.py
├── denoise.py
├── SampleGenerate.py
├── SampleView.py
└── requirements.txt

---

Training Results

Generator and Discriminator Loss

---

Spectrogram Comparison

Real Spectrograms

Generated Spectrograms

---

Generated Audio Samples

Generated audio outputs produced by the trained SpecGAN model are available in the generated_audio directory.

Examples:

• generated_clean_denoised.wav
• generated_clean_denoised_eq.wav

These samples demonstrate the model's ability to synthesize realistic orca vocalization patterns from generated spectrograms.

---

Installation

git clone https://github.com/7mgppp/Spectrogram-Based-GANs-for-Orca-Call-Synthesis-I.git

cd Spectrogram-Based-GANs-for-Orca-Call-Synthesis-I

pip install -r requirements.txt

---

Training

python train.py

---

Generate Samples

python SampleGenerate.py

---

View Generated Samples

python SampleView.py

---

Technologies Used

• Python
• PyTorch
• NumPy
• Librosa
• Matplotlib
• Scikit-learn

---

Results Summary

• Learned temporal and harmonic structures of orca calls.
• Generated realistic spectrogram representations.
• Achieved an FID score of 46.3.
• Successfully reconstructed synthesized audio samples.

---

Future Work

• Improve FID scores.
• Train on larger bioacoustic datasets.
• Explore diffusion-based audio generation.
• Improve waveform reconstruction quality.
• Investigate conditional audio generation.

---

Applications

• Marine bioacoustics research
• Wildlife conservation
• Acoustic data augmentation
• Animal communication studies
• Generative audio modeling

---

Author

Miilee Sharma

B.Tech Computer Science (Data Science)
Manipal University Jaipur

GitHub: https://github.com/7mgppp

---

License

This project is released under the MIT License.