Abstract
Generative Adversarial Networks (GAN) have seen great improvement in modelling synthetic data with results unparalleled in the field of deep learning. This research explores data augmentation for cover song identification by extending state-of-the-art GAN frameworks used in image processing to cover song identification. Research in this domain is still in the early phases, this work explores the implementation of a GAN model capable of generating new features for cover song identification. The implementation utilises the Deep Convolutional Generative Adversarial Networks (DCGAN) to learn the underlying distribution of the sample dataset provided by Breathe Music, the conditional Generative Adversarial Networks (cGANs) was used for the encoding of the class to give control over which class in the distribution is generated. A control model was also created using an industry-standard dataset, CIFAR-100, to validate the approach. The qualitative and quantitative results were obtained for both two datasets, and a comparative analysis of both models was undertaken. The model results demonstrated a strong performance in capturing the underlying data distribution for the CIFAR100 dataset but performed poorly for the custom music dataset provided by Breathe Music using the statistical probability distribution plot, and Fr´echet Inception Score (FID) metric.