Speech Bandwidth Extension Based on Wasserstein Generative Adversarial Network

Artificial bandwidth extension (ABE) algorithms have been developed to improve the quality of narrowband calls before devices are upgraded to wideband calls. Most methods use deep neural networks (DNN) to establish a nonlinear relationship between narrowband components and wideband components, so as to predict high frequency components from narrowband. Although traditional convolutional deep neural networks (CNN) trained on the minimum mean square error (MSE) can bring high peak Signal to Noise Ratio (SNR), it usually lacks high frequency details and has poor generalization. In this paper, we propose a speech signal super-resolution Wasserstein generative adversarial network (SRWGAN). In this paper, we propose a new speech signal super resolution method based on Wasserstein generated adversarial network, where a network joined with adversarial learning is designed and a perceptual loss function including adversarial loss and regression loss is derived. The simulation results shows that the proposed scheme is better than the traditional minimum mean square error training network in predicting high frequency components.