多通道语音增强算法研究

发布时间：2018-06-10 18:43

本文选题：麦克风阵列 + 卡尔曼滤波　；参考：《湖北工业大学》2017年硕士论文

【摘要】：在生活环境中,语音信息都会受到周围噪声的干扰,从而影响许多语音设备的性能。麦克风阵列具有空间选择特性与高信号增益特性,从而可以获得更多的有效语音信息,这样就可以把噪声更好的从语音中分离出来。但是仅仅只用单一的麦克风阵列算法也难满足要求,近年来,基于麦克风阵列的语音增强方法和其它算法相结合的方法己经逐渐成为语音增强技术的研究热点。本论文也着手研究基于麦克风阵列的结合算法的语音增强算法。主要工作如下:针对传统的卡尔曼滤波中噪声谱估计问题,提出一种基于能量直方图噪声功率谱估计的改进方法。其基本方法是基于过去功率谱值的直方图来得到噪声谱估计,即对新进的每一帧,首先在几百毫秒的时间窗上构建一个功率谱值的直方图。利用直方图中最大值所对应的的功率值代替噪声功率谱值。通过两个仿真实验来验证算法的消噪性能。实验一是最小统计跟踪的噪声谱估计算法与基于能量数量值的噪声谱估计的改进算法的噪声谱估计仿真。仿真结果显示改进的噪声谱估计方法更接近真实的噪声。实验二是把改进噪声谱估计算法引入到卡尔曼滤波中,对改进的卡尔曼滤波算法滤波算法进行仿真验证。仿真图显示信号的语音成分保留下来,没有造成语音过估计。但改进的卡尔曼滤波的还是存留较多的背景噪声,这是因为单一通道的语音增强算法并不能满足要求,需要利用多麦克风阵列获得更多的语音信息完成去噪。针对改进的卡尔曼滤波存在的问题,以及传统麦克风阵列语音增强算法存在的不足,本文提出了将改进的卡尔曼滤波与广义旁瓣相消器相结合的方法。利用改进卡尔曼滤波作为后置滤波,克服了GSC对非相干噪声的抑制不明显以及残留较多的噪声的缺点。通过仿真实验比较几种算法,结果显示相对于传统的GSC自适应算法和后置滤波算法,改进算法滤波后的背景噪声更小,具有更高的输出信噪比。同时采取主观和客观的评价机制对几种算法进行比较,都证明了基于GSC结构的改进算法的消噪性能更好。
[Abstract]:In the living environment, the speech information will be disturbed by the surrounding noise, which will affect the performance of many speech devices. The microphone array has the characteristics of spatial selection and high signal gain, so that more effective speech information can be obtained, and the noise can be better separated from the speech. However, only a single microphone array algorithm is difficult to meet the requirements. In recent years, speech enhancement methods based on microphone array and other algorithms have gradually become the focus of research on speech enhancement technology. This thesis also studies the speech enhancement algorithm based on microphone array. The main work is as follows: aiming at the problem of noise spectrum estimation in traditional Kalman filtering, an improved method based on energy histogram noise power spectrum estimation is proposed. The basic method is to estimate the noise spectrum based on the histogram of the past power spectrum, that is, to construct a histogram of the power spectrum at the time window of several hundred milliseconds for every new frame. The noise power spectrum is replaced by the power value corresponding to the maximum value in the histogram. Two simulation experiments are carried out to verify the performance of the algorithm. The first experiment is the simulation of the noise spectrum estimation algorithm based on the minimum statistical tracking and the improved noise spectrum estimation algorithm based on the energy quantity. Simulation results show that the improved noise spectrum estimation method is closer to the real noise. In the second experiment, the improved noise spectrum estimation algorithm is introduced into the Kalman filter, and the improved Kalman filtering algorithm is simulated and verified. The simulation diagram shows that the speech component of the signal is preserved, and the speech is not overestimated. But the improved Kalman filter still has more background noise, which is because the single channel speech enhancement algorithm can not meet the requirements, and needs to use multi-microphone array to obtain more speech information to complete the de-noising. Aiming at the problems of the improved Kalman filter and the shortcomings of the traditional microphone array speech enhancement algorithm, this paper proposes a method to combine the improved Kalman filter with the generalized sidelobe canceller. The improved Kalman filter is used as the post-filter to overcome the shortcomings of the GSC in suppressing incoherent noise which is not obvious and that there is more residual noise. The simulation results show that compared with the traditional GSC adaptive algorithm and post-filtering algorithm, the improved algorithm has less background noise and higher output signal-to-noise ratio than the traditional GSC adaptive algorithm and post-filter algorithm. At the same time, the subjective and objective evaluation mechanisms are used to compare several algorithms. It is proved that the improved algorithm based on GSC structure has better performance in noise reduction.
【学位授予单位】：湖北工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.35

【相似文献】