【作者】 应冬文； 颜永红； 付强； 国雁萌；

【Author】 Dongwen YING,Yonghong YAN,Qiang FU,Yanmeng GUO (Institute of Acoustics,Chinese Academic of Sciences,Beijing 100190,China )

【机构】 中国科学院声学研究所；

【摘要】 噪声功率谱估计是消噪算法不可或缺的组成部分。本文提出了一种算法,它基于约束的高斯混合模型在对数谱域追踪噪声功率谱,这一模型分别由语音和非语音两个高斯分量组成,这里的非语音均值即是噪声对数功率谱的无偏估计。本算法的焦点问题是估计高斯混合模型的参数集,首先通过EM算法离线初始化模型,然后随着输入信号的变化,通过一阶回归平均逐帧更新模型参数。为了模型的稳定性,我们引入了一些约束到模型中。本算法和主流的算法,如最小统计(MS)、最小控制的回归平均(MCRA)、提高版本的最小控制的回归平均(IMCRA),进行了比较,实验表明本算法优于这些主流的算法。更多还原

【Abstract】 Noise power estimation is an indispensible component to noise reduction algorithms.In this paper,we propose a constrained Gaussian mixture model to track noise power in the log-spectral domain.This model comprises a speech and nonspeech components,where the nonspeech mean is the unbiased estimation of the noise power.The focus of this algorithm is on estimating the GMM parameter set.This parameter set is firstly estimated by EM algorithms,and then is incrementally updated by a first-order recursive averaging to follow signal variation.For the sake of reliability, some constraints are introduced to this model.The proposed algorithm is compared with the conventional algorithms such as Minimum statistics,MCRA and IMCRA.The experiment results show its promising performance.更多还原