N2N2N: A Clean Data Independent Speech Enhancement Approach with Modified cGAN

This work presents a novel approach for speech enhancement in scenarios where clean data is unavailable but noise type is known, using a modified Conditional Generative Adversarial Network (cGAN) with a Residual Attention UNet (RAUNet) as the generator. The proposed Noise-to-Noisy-to-Noise (N2N2N) algorithm works on known noise types to iteratively refine speech signals, achieving high performance even in low signal-to-noise ratio (SNR) conditions such as -15 dB. We demonstrate that our method surpasses existing solutions in terms of robustness, phase preservation, and computational efficiency. Experimental results validate the effectiveness of our approach, showcasing significant improvements in PESQ, STOI, CBAK, COVL and SSNR metrics across various noise environments.