Integrating Non-Matching Audio Distance Loss to the CMGAN Speech Enhancement

The Conformer-based Metric Generative Adversarial Network is a well-known speech enhancement architecture that successfully removes noise from both the magnitude and complex spectrograms of input speech. In CMGAN, multiple aspects of loss are considered to learn both the generator and discriminator. In this study, we propose revising the loss function by further considering NOMAD loss, an audio metric that can be applied to any non-matching reference that is perceptually differentiable. The experiments using the VoiceBank-DEMAND dataset show that incorporating NOMAD loss during CMGAN training results in a significant improvement in PESQ and STOI, two popular objective speech enhancement metrics.