Multi-channel Speech Enhancement with Multiple-target GANs
Jing Yuan, Changchun Bao · 2020
In noisy scenes, speech enhancement is an important technology to improve the speech quality. In this paper, a multi-channel speech enhancement algorithm with multiple-target Generative Adversarial Networks (GANs) is proposed. Firstly, using the spatial characteristics of microphone array, the mask of target speech signal is generated by the multiple-target GAN (MT-GAN). Secondly, the mask is estimated based on complex Gaussian mixture model (CGMM), which is combined with the mask predicted by network in an iterative way to obtain a more robust speech enhancement system. Finally, the estimated mask is used to construct beamformer. Thus, the noisy speech is enhanced by the constructed beamformer. The experimental results show that compared with the reference methods, the speech quality and intelligibility of the proposed method are improved effectively.