"CMGAN: Conformer-Based Metric-GAN for Monaural Speech Enhancement" presented by Dr. Sherif Abdulatif. Based on the IEEE Xplore® article under the same title. Published: IEEE Transactions on Audio, Speech, and Language Processing, April 2024. Visit article article here.
Modern speech enhancement systems must operate reliably under a wide range of real-world distortions, including background noise, reverberation, and bandwidth limitations. This webinar presents CMGAN, a Conformer-Based Metric-GAN framework for monaural speech enhancement in the complex time-frequency domain.
The talk will introduce the motivation behind combining magnitude masking, complex spectrogram refinement, conformer-based sequence modeling, and metric-guided adversarial learning. CMGAN uses a shared encoder to process magnitude and complex spectral components, followed by two-stage conformer blocks that capture temporal and frequency dependencies. A dedicated mask decoder enhances the magnitude representation, while a complex decoder refines the real and imaginary components to improve both magnitude and phase reconstruction. In addition, a metric discriminator is used to optimize perceptually meaningful quality measures that are not directly differentiable.
The webinar will discuss how this architecture addresses three major speech enhancement tasks: speech denoising, dereverberation, and speech super-resolution. It will also summarize key experimental findings, including comparisons with state-of-the-art methods, ablation studies on architectural and loss-function choices, and evaluations using objective metrics, DNS-MOS, and listening tests. The presentation aims to provide both conceptual insight and practical guidance for researchers interested in deep learning-based speech enhancement, complex time-frequency modeling, and perceptual optimization.
