一种改进的语音谱减去噪算法  

Improved speech denoising method based on spectral subtraction

作　　者：冯纪强[1] 徐晨[1] 张维强[1] 

机构地区：[1]深圳大学智能计算科学研究所,广东深圳518060

出　　处：《计算机工程与应用》2007年第23期80-82,共3页Computer Engineering and Applications

基　　金：广东省自然科学基金(the Natural Science Foundation of Guangdong Province of China under Grant No.06105776);广东省科技计划项目(No.2005B10201009);深圳市科技计划费用项目(No.200511)

摘　　要：在谱减去噪过程中噪音的功率谱估计一般根据经验而定,对此提出了一种改进方法,利用含噪语音的短时能零积和基本谱减法,得到语音起止点和噪音功率谱估计,有利于在不同语音阶段对含噪语音进行谱减去噪。最后利用无音阶段噪音特点对去噪之后的残留噪音进行残差处理以彻底去除噪音。仿真实验表明该方法比传统单一的谱减去噪方法效果理想。In the course of de-noising using spectrum subtraction,power spectra of noise is generally decided via traditional experience.A new approach is proposed to change the traditional approach.Speech start-end points and power spectra of noise are firstly achieved through general spectral subtraction methods and little-time energy-zero-product.Then further denoising through spectral subtraction methods is more effective during the different respective speech segment.Residual noise reduction is proposed based on the characteristic of noise in the non-speech activity.When this method is applied to denoise speech corrupted by additive white Gaussian noise,simulation experiment results illustrate that this method is more effective than the traditional spectral subtraction methods.

关 键 词：语音信号 谱减法 短时能零积 残差处理 去噪 

分 类 号：TN911[电子电信—通信与信息系统]