EnvGAN: a GAN-based augmentation to improve environmental sound classification

Several deep learning algorithms have emerged for the automatic classification of environmental sounds. However, the non-availability of adequate labeled data for training limits the performance of these algorithms. Data augmentation is an appropriate solution to this problem. Generative Adversarial Networks (GANs) can successfully generate synthetic speech and sounds of musical instruments for classification applications. In this paper, we present a method for GAN-based augmentation in the context of environmental sound classification. We introduce an architecture named EnvGAN for the adversarial generation of environmental sounds. To validate the quality of the generated sounds, we have conducted subjective and objective evaluations. The results indicate that EnvGAN can produce samples of various domains with an acceptable target quality. We applied this augmentation technique on three benchmark ESC datasets (ESC-10, UrbanSound8K, and TUT Urban Acoustic Scenes development dataset) and used it for training a CNN-based classifier. Experimental results show that this new augmentation method can outperform a baseline method with no augmentation by a relatively wide margin (10-12% on ESC-10, 5-7% on UrbanSound8K, and 4-5% on TUT). In particular, the GAN-based approach reduces the confusion between all pairs of classes on UrbanSound8K. That is, the proposed method is especially suitable for handling class-imbalanced datasets.