Generative Adversarial Networks in Speech Enhancement: A Survey

Generative adversarial networks are a powerful type of model in deep learning. They have been successfully applied within different domains. This review focuses on the usage of generative adversarial networks for speech enhancement. In total, 87 studies are analyzed and summarized, with their publication period ranging from the earliest attempts to papers released in November 2025. This survey aims to provide the necessary background information for researchers planning to use or already applying generative adversarial networks to enhance speech signals. It examines generative adversarial network-based models by analyzing signal representations at the input and output, network architectures, and loss function formulations. Temporal trends in these design decisions are analyzed to illustrate the evolution of the models over time. The surveyed models are further compared based on their performance results on standard benchmark dataset. This evaluation helps identify which models achieve the best performance for specific speech enhancement tasks addressed in the literature. The limitations and future research directions reported in the surveyed studies are summarized, along with additional insights derived from model analysis.