Marine Audio Augmentation and Validation Framework Using WaveGAN

Project Synopsis & Overview

This project develops a comprehensive audio data augmentation framework for marine species classification, particularly whale vocalizations. The system collects real marine audio recordings and performs preprocessing operations such as downsampling, denoising, segmentation, silence removal, and normalization. A WaveGAN deep learning model is trained using the processed audio data to generate realistic synthetic marine audio samples. Traditional augmentation techniques, including time-stretching, pitch-shifting, noise injection, and random gain adjustment, are also applied to increase dataset diversity. The generated and augmented audio samples are combined with the original dataset to create an enhanced dataset for marine bioacoustic applications. The quality of the generated audio is validated using waveform and spectrogram visualization and quantitatively using Kullback–Leibler divergence. The framework aims to improve the diversity, balance, and robustness of datasets used for marine species identification and monitoring.

Complete Technical Specifications