Abstract: In order to get a better performance in processing the speech which changes rapidly，an improved speech enhancement algorithm based on self-adaptive averaging factor method is proposed in this paper.Simulation with Itakura-Saito distance，speech wave picture，spectral picture Experimental results demonstrate that the proposed algorithm has better performance of speech articulation without any perceptional musicality，compared to another modified spectral subtraction algorithms under the same level of noise reduction，and this superiority is more significant at very low SNR.

Key words: speech enhancement, spectral subtraction of geometric approach, Minimum Mean Square Error（MMSE）, Itakura-Saito distance

摘要： 针对几何谱减算法在处理快速变化的语音时产生语音畸变的缺点，提出一种基于最小均方误差算法估计每帧语音信号的每一个频率分量上的平滑系数，产生自适应帧频率分量平滑系数代替固定值的平滑系数来估计先验信噪比，从而得到更加接近于真实情况的先验信噪比。通过计算板仓-斋藤距离，及利用仿真波形图、语谱图对算法进行客观测试，结果表明新算法相对其他谱减法在相同的去噪度下，语音畸变度最小且几乎察觉不到音乐噪声；特别是在低信噪比非平稳环境下，相对其他谱减法的优势更加显著。

关键词: 语音增强, 几何谱减法, 最小均方误差, 板仓-斋藤距离