基于条件生成对抗网络的语音增强

Speech Enhancement Based on Conditional Generative Adversarial Nets

语音增强技术目前有传统的方法和深层神经网络的方法,传统的方法假设语音信号和噪音之间的关系做出假设,这往往造成语音增强后效果不佳,深层神经网络的方法具有根据数据进行去噪的能力,避免了前期的假设。论文使用条件生成式对抗网络(Conditional Generative Adversarial Nets)的方法对语音进行增强,采用有监督的训练方式,在训练过程中加入带噪语音信号,能够有效地指导训练的进行。采用PESQ对增强后的语音质量进行评价,实验结果显示,论文方法能够有效地对带噪语音进行增强。

At present,speech enhancement technology has the traditional method and the deep neural network method.The traditional method assumes the relationship between speech signal and noise,which often leads to poor effect after speech enhance⁃ment.The deep neural network method has the ability to remove noise according to data,avoiding previous assumptions.In this pa⁃per,the method of conditional generative adversarial nets is used to enhance the speech,and a supervised training method is adopt⁃ed.Adding noisy speech signals to the training process can effectively guide the training.PESQ is used to evaluate the speech quali⁃ty after enhancement.Experimental results show that this method can effectively enhance the speech with noise.