多声道语音增强系统中维纳滤波器的研究及其在语音识别中的应用

Study of Wiener Filter in Multichannel Noise Suppression Systems and Its Application in Speech Recognition

本文讨论了多声道噪声抑制系统，提出了一种新的采用维纳滤波器消噪的组合法，其消噪性能优于Zelinski法[12]的消噪性能。作者根据人耳的听觉特性修改了以往的信噪比（SNR）失真测度而追求增强语音谱与纯净语音谱之间的吻合，并提出了维纳滤波器转移函数的修正估计法以求获得更佳的谱一致性。最后，就各种消噪方法对语音识别系统的影响作了大量的试验。实验结果表明，消噪系统性能的提高有效地改善了识别系统的正识率。

This paper studies multichannel noise suppression systems from different point of views to improve the performance of speech enhancement systems. In the first part of this paper, we propose a new combination algorithm, both its SNR and the intelligibility of the filtered speech are better than those of Zelinski's method[12].In th second part we utilize the characteristic of human ear and suggest to modify the distance measure of distortion instead of using the SNR. There are two modifications of estimation of power spectrum of speech and the transfer function of Wiener filter in order to obtain a bette speech spectrum. The third part discusses the application of our various algorithms to noisy speech recognition systems