Adversarial Training Using Inter/Intra-Attention Architecture for Speech Enhancement Network

In this paper, we propose a new adversarial training for the end-to-end speech enhancement network. Taking the advantage of getting the paired training waveform, a new attention module is introduced into the proposed discriminator to incorporate the information of the desired waveform. Since this attention module has a role of the inter- and intra-attention mechanism, it helps the discriminator to distinctly distinguish the structural features underlying in the desired waveform and the waveform generated by the speech enhancement network. Unlike the other conditional generative adversarial networks, the proposed training architecture can simultaneously minimize the adversarial loss and the distortion loss. Through the simulation experiments for speech enhancement, we reveal that the proposed adversarial training can provide a significant performance.