基于子空间域噪声特征值估计的语音增强方法  

Speech Enhancement Based on Noise Eigenvalue Estimation in Subspace Domain

作　　者：吴北平[1] 李辉[1] 戴蓓倩[1] 陆伟[1] 

机构地区：[1]中国科学技术大学多媒体计算与通信教育部-微软重点实验室,合肥230026

出　　处：《信号处理》2009年第3期460-463,共4页Journal of Signal Processing

摘　　要：传统信号子空间语音增强方法中均采用语音活动检测方法估计噪声,当噪声统计特性发生变化或信噪比较低时使用语音活动检测不能准确的估计噪声。本文给出一种基于子空间域噪声特征值估计的语音增强方法。结合语音存在概率对带噪语音协方差矩阵在每个特征向量上的特征值递归平滑得到噪声估计,可以在每一帧内更新噪声特征值。该方法不需要区分有声段和无声段,能够更加准确的反映当前时刻的噪声水平,具有鲁棒性。实验结果表明本文方法要优于传统的子空间语音增强方法。Conventional subspace speech enhancement methods used voice activity detection to estimate noise. When the statistics of noise are changing or signal-noise-ratio （SNR） is low, the noise estimated value by voice activity detection is not exact. We propose a new approach for enhancement of noisy speech based on noise eigenvalue estimation in subspace domain. In contrast to other methods, our approach does not use accurate voice activity detection. Instead it recursive smoothes past noisy signal eigenvalues without any distinction between speech activity and speech pause, using a smoothing parameter that is adjusted by the speech presence probability in eigenvectors. Our method updates the noise estimate within every frame, and estimates noise in eigenvectors more exactly, and then improves the quality of the enhanced speech subsequently. Through many simulations, it is demonstrated that the proposed approach outperforms the conventional subspace method.

关 键 词：噪声估计 信号子空间 语音增强 

分 类 号：TN912.35[电子电信—通信与信息系统]