基于Householder变换的声学降噪理论研究

Noise reduction consists of recovering clean speech signal from the noisy signals, which are corrupted by unwanted additive noise. It can find numerous applications in various fields of speech processing, such as teleconferencing, immersive communications, hearing aids systems and so on. As the rapid development of the acoustic industry, it has a high demand for noise reduction algorithms with excellent performance, high robustness and practical value. However, the performance of traditional noise reduction methods too much depends on the noise estimation accuracy and voice activity detection process. As a result, we intend to establish a comprehensive novel theory framework of noise reduction based on Householder transformation. Householder transformation converts the noisy signal coefficients into two components: speech dominated and noise dominated component, and then optimal filters can be obtained. One of the major advantages of this approach as compared to many traditional noise reduction techniques is that it does not need a noise estimator or a voice activity detector. Our method can also be combined with some other acoustic signal processing algorithms based on Householder transformation to form a complete theoretical system, which will be helpful for the innovation and development of domestic acoustic industry.

声学降噪，就是从受噪声污染的语音信号中提取出纯净语音的过程，它被广泛应用于远程会议、临境通讯、助听设备等领域。随着声学产业的飞速发展，急需与之相匹配的性能优异、鲁棒性高、实用性强的声学降噪算法。但是，传统声学降噪理论过于依赖噪声估计精度和语音活动检测过程，限制了降噪算法的使用性能。在这样的背景下，本项目拟建立一套完善的基于Householder变换的声学降噪理论方案。利用Householder变换，我们将信号转换为两个部分：第一部分主要由纯净信号构成，第二部分主要由噪声构成，利用分解后的结果进一步推导相应滤波器。与传统声学降噪过程比较，本项目所提出的方法不需要进行噪声估计或者语音活动检测，为声学降噪理论的发展提供了新的思路。另外该方法还可与基于Householder变换的其他声信号处理算法相结合，形成一套完备的理论体系。这些研究结果将有力推动国内声学产业的创新和发展。

本研究项目以基于Householder变换的声学降噪方法作为出发点，进一步提出基于Householder变换的声源定位理论，最终通过完善深化，形成了一套完备的基于几何投影变换的多通道声学信号处理理论体系。本项目的主要创新性和突破性成果体现在以下三个方面：1、提出了基于Householder变换的声学降噪算法，该方法区别于传统方法的最大优势在于其不需要进行噪声估计；2、提出了基于Householder变换的声源定位算法，该方法对原始多通道信号进行分解，利用分解后的结果构造定位代价函数，并搜索区域内代价函数的峰值位置作为声源位置的估计。该方法提供了定位理论的新思路，并且定位准确，灵活度较高；3、在以上理论的基础之上，本研究系统性提出一套完整的基于几何投影变换的多通道声学信号处理理论。在此理论框架下，我们构造了基于几何投影变换的四类代价函数和三类宽带融合方法，并详细解释了传统定位方法的几何意义。这一理论直观解释了定位算法的本质，并且找到了各种算法之间的联系和等价性，从而帮助我们对各算法的优缺点进行把握，为寻求性能更加优秀的定位算法提供了理论指导。