采用复高斯分布模型的两步噪声幅度谱估计算法

摘要

The estimate for the amplitude of noise signal plays an important role in many noise reduction or speech enhancement methods.However,compared with the noise power estimation,less attention has been paid to the amplitude estimation in the past years.In addition,the frequently-used voice activity detection (VAD) algorithm estimates or updates the noise amplitude during only speech absence area,which leads to an inferior performance in non-stationary noise conditions.To overcome such drawback,tow novel noise amplitude estimators working with two steps are proposed in this paper based on the assumption of complex Gaussian model.The estimation of noise power is achieved by soft decision (SD) method in the first step,and then the indirect estimators are subsequently obtained by using the relationship between the power and amplitude under the complex Gaussian distribution.The results of simulations indicated that the presented estimators can lead to significantly better speech quality than the frequently-used VAD method under various noise conditions.%噪声幅度谱估计是有效抑制外界噪声干扰、提高语音增强算法整体输出性能的重要环节.但目前针对该问题的研究相对较少,常用的语音激活检测算法只能在语音不存在阶段对噪声信号的幅度谱进行更新或估计,无法适用于更为复杂的非平稳噪声环境.为克服这一问题,本文基于噪声频谱的复高斯分布模型假设,提出了新型的两步噪声幅度谱估计算法.算法首先采用软判决技术计算噪声信号的功率谱,然后再结合复高斯分布条件下信号幅度谱和功率谱之间的数学关系间接地获取噪声幅度谱的估计.文中基于这一结论给出了两种估计算法,并在多种噪声环境下对它们的性能进行了仿真评估,其测试结果有效表明了提出算法优良的估计性能.