Bridging the Gap: GANs as a Solution for Data-Scarce Industrial Audio Classification

Data scarcity remains challenging when training deep learning models. This is particularly true in audio classification tasks within the Industrial Sound Analysis (ISA) domain, where the collection of training data represents a significant investment. In this work, we investigate the usage of Generative Adversarial Networks (GANs) in generating synthetic data to improve model performance in low-resource domains. An audio classification model is then trained on a combination of this synthetic data with (comparatively few) real-world examples, and evaluated on unseen real data. We view this approach as analogous to data augmentation, where instead of transforming existing data, the GAN generates novel data for training diversification. To explore the method's potential for various use cases, we run experiments on two different ISA datasets, each exhibiting different audio characteristics, imposing various levels of training data scarcity on each. We show that our method leads to a significant increase in classification accuracy for one of the two datasets, and we analyze the factors which make it successful.