一种基于Laplacian语音模型的语音增强算法  

Speech Enhancement With Laplacian Speech Modeling

作　　者：邹霞[1] 吴其前[1] 张雄伟[1] 

机构地区：[1]解放军理工大学 通信工程学院,南京210007

出　　处：《信号处理》2007年第2期195-199,共5页Journal of Signal Processing

基　　金：江苏省自然科学基金(BK2006001);江苏省图像处理与图像通信实验室资助项目(KJS03036)

摘　　要：本文提出了一种新的基于Laplacian语音模型的语音增强算法。首先,在假定语音和噪声的短时DCT系数分别服从Laplacian和Gaussian分布的基础上,推导了最小均方误差意义下的语音信号短时DCT系数估计;然后,根据语音存在概率估计,提出了语音信号短时DCT系数估计的修正因子。在增强算法中,提出了面向判决的Laplacian语音模型参数估计和基于Laplacian语音模型的改进最小量控制递归平均(IMCRA)噪声估计算法。仿真结果表明,本文算法不仅在噪声抑制性能方面优于近两年国际上提出的几种基于Gaussian语音模型的语音增强算法,而且在增强语音质量方面也具有更好的性能。In this paper, a new speech enhancement system with Laplacian speech modeling is proposed. The clean speech components are estimated by Minimum-Mean-Square-Error （MMSE） estimator under the assumption that the clean speech DCT coefficients are modeled by a Laplacian distribution and the noise DCT coefficients are modeled by a Gaussian distribution. Then, the MMSE estimator under speech presence uncertainty and a Laplacian model is derived. Furthermore, the proper estimators of the statistical parameters are proposed. The speech Laplacian parameter is estimated by a decision-directed method and the noise power is estimated by the IMCRA algorithm with Laplacian model. The simulation results show that the proposed algorithm yields less residual noise and better speech quality than the Gaussian based speech enhancement algorithms proposed in recent years.

关 键 词：语音增强 语音模型 最小均方误差 噪声估计 

分 类 号：TN912.35[电子电信—通信与信息系统]