Speech Denoising Based on Estimating Sound Source Direction for Car Navigation System

我们针对汽车导航系统等小型设备的语音去噪方法进行了研究和开发。我们的方法是基于1)估计声源方向和2)频谱减法。我们开发了一种从两个麦克风接收的信号中估计多个声源方向的方法。该方法利用语音信号的谐波结构对其频谱分量进行分类，从而实现声源方向的估计。采用矢量量化技术对频谱分量进行分类。以两种声源（扬声器）为对象进行了仿真评价实验。仿真结果表明，在精度小于5度的情况下，该方法的估计正确率达到83%。2)提出了改进的双传声器频谱减法。该方法可以有效地降低道路噪声等非平稳噪声。我们在数字信号处理器（C6713DSK）上进行了该方法的实现实验。

We carried out research and development regarding a speech denoising method for a small device such as a car navigation system. Our method has been developed based on 1) estimating sound source direction and 2) spectrum subtraction.1) We have developed a method of estimating multiple sound source directions from signals received by two microphones. In this method, estimation of sound source directions is performed by classifying spectrum components of speech signal using its harmonic structure. The technique of vector quantization is used for the classification of spectrum components. The evaluation experiment for our method by the simulation was performed about two sound sources (speakers).The simulation results reveal that the proposed method attained 83% of rate of a correct estimation int the accuracy of less than 5 degree.2) We proposed the improved spectrum subtraction method with two microphones. This method can reduce unstationary noise such as the road noise. We carried out implementation experiment of this method on the digital signal processor (C6713DSK).