A two-stage complex network using cycle-consistent generative adversarial networks for speech enhancement

Cycle-consistent generative adversarial networks (CycleGAN) have shown their promising performance for speech enhancement (SE), while one intractable shortcoming of these CycleGAN-based SE systems is that the noise components propagate throughout cycle and cannot be completely eliminated. Additionally, conventional only estimate spectral magnitude, phase unaltered. Motivated by multi-stage learning concept, we propose a novel two-stage denoising system combines magnitude enhancing network subsequent complex refining in this paper. Specifically, first stage, model responsible estimating which subsequently coupled with original noisy to obtain coarsely enhanced spectrum. After that, second stage applied further suppress residual clean mapping network, pure complex-valued composed 2D convolution/deconvolution temporal-frequency attention blocks. Experimental results on two public datasets demonstrate proposed approach consistently surpasses previous one-stage CycleGANs other state-of-the-art terms various evaluation metrics, especially background suppression. • We network. decompose spectrum estimation into sub-tasks, i.e., maginitude phase. The outperforms many approaches.