Speech Dereverberation Based on Improved Wasserstein Generative Adversarial Networks
Lufang Rao, Junmei Yang · Journal of Physics Conference Series · 2020
Abstract In reality, the sound we hear is not only disturbed by noise, but also the reverberant, whose effects are rarely taken into account. Recently, deep learning has shown great advantages in speech signal processing. But among the existing dereverberation approaches, very few methods apply deep learning at the waveform level. In addition, in the case of sever reverberation, the conventional speech dereverberation methods perform poorly, such as MCLP (multi-channel linear prediction). We proposed a new speech dereverberation method in this paper, which is based on improved WGAN (Wasserstein Generative Adversarial Networks), called WGAN-GP, whose generator uses strided-convolutional networks and the discriminator is structured on DNNs. Due to the addition of the gradient penalty item, WGAN-GP improves the stability of training and the generalization of the model. In the case of severe reverberation, according to the experimental results, the proposed system can perform better than MCLP. As the proposed method based on WAGN-GP can improve speech quality, speech signal processing systems may be able to apply it to pre-processing stage.