结合波束形成和GAN网络的多通道语音增强研究

Multi-channel Speech Enhancement based on Beamforming and GAN Network

后端滤波处理是多通道语音增强系统中一种比较常用的技术,其目的是为了进一步提高语音增强系统的性能,提高波束形成后的输出信噪比。但是,常用的后滤波方法需要相当繁琐的参数调整过程才能实现噪声抑制和语音质量之间的合理权衡。本文提出一种基于最小方差无畸变(MVDR)波束形成和生成对抗深层神经网络相结合的多通道语音增强算法。前端使用波束形成器对信号进行初步增强;后端滤波处理采用生成对抗深层神经网络,避免了繁琐的参数调整过程。实验系统是通过MATLAB和Tensor Flow仿真实现,结果证明了该方法的有效性。

Post filtering process is a common technique in multi-channel speech enhancement system. Its purpose is to further improve the performance of speech enhancement system and improve the output signal-to-noise ratio after beamforming. In order to realize the reasonable tradeoff between noise suppression and speech quality, the commonly used post-filtering methods require a rather cumbersome process of parameter adjustment. In this paper, a new multi-channel speech enhancement algorithm combined beamforming method based on minimum variance distortionless（MVDR） and generative adversarial neural networks（GAN） is proposed. The beamformer is used in the front end to preliminary enhance the signal. The back-end filter use the proposed GAN to enhance the speech signal, which avoids the complicated parameter adjustment process. The experimental system is realized by Matlab and Tensor Flow simulation. The results show that the method is effective.