基于噪声短时谱动态估计的语音增强谱减算法  

Spectral subtraction algorithm for speech enhancement based on dynamic estimation of noise short-time spectrum

作　　者：李银国[1] 薛雯[1] 徐洋[1] 

机构地区：[1]重庆邮电大学自动化学院,重庆400065

出　　处：《重庆邮电大学学报（自然科学版）》2010年第2期127-130,共4页Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)

基　　金：重庆市科委重大科技攻关项目(2008AA6002);电子信息产业化发展基金(工信部财[2008]355号)~~

摘　　要：针对谱减法增强语音后残留"音乐噪声"明显,影响语音清晰度的问题,在分析现有的噪声谱减算法基本特性的基础上,提出一种基于噪声短时谱动态估计的语音降噪方法。通过估计噪声短时功率谱及其变化趋势,动态调整谱减法中的过减系数,有效抑制残留噪声。仿真实验表明,新的谱减算法在提高语音信噪比的同时保持较小的失真度,具有较好的增强效果。In order to reduce musical noise in spectral subtraction method and improve the intelligibility of enhanced speech, the basic characteristics of exiting spectral subtraction were analyzed. A noise-robust method based on dynamic estimation of noise short-time spectrum was proposed. Over-subtraction of spectral subtraction could be modified dynamically to suppress residual noise effectively by estimating noise short-time power spectrum and its trends. Experimental results show that the new spectral subtraction algorithm improves the speech signal to noise ratio while maintaining a smaller distortion and it has good enhancement effect.

关 键 词：语音识别 语音增强算法 噪声短时谱估计 谱减法 过减系数 

分 类 号：TN912.34[电子电信—通信与信息系统]