基于连续噪声谱估计的谱减法语音增强算法

Spectral Subtraction Speech Enhancement Algorithm based on Continuous Noise Spectral Estimation

传统谱减算法简单高效,但是其端点检测不准确,噪声估计仅仅采用"寂静段"的统计平均,会引入起伏较大的"音乐噪声",影响通信效果。针对传统端点检测和噪声估计不准确的问题,提出了一种基于连续噪声谱估计的谱减法语音增强算法,通过对噪声谱进行不间断更新来准确定位语音端点和噪声估计,然后在此基础上进行功率谱过减及半波整流,运用维纳滤波器对语音信号进行平滑处理。结果表明,改进的谱减法能够消除背景噪声,并有效滤除"音乐噪声",得到更好的语音可懂度和清晰度。

Traditional spectral subtraction algorithm is simple and efficient, and however its endpoint detection is inaccurate, and only the statistical average of ＂silent segments＂ is adopted for noise estimation, this would cause ＂music noise＂ with large fluctuations, and directly affect the communication effect. Aiming at the problem of inaccuracy for traditional endpoint detection and noise estimation, a spectral subtraction speech enhancement algorithm based on continuous noise spectrum estimation is proposed. The noise spectrum is uninterruptedly updated, so as to accurately locate the speech endpoints and perform noise estimation. On this basis, power spectrum over-reduction and half-wave rectification are done, while the Wiener filter is used to smooth the speech signal. The experiment indicates that the modified spectral subtraction could eliminate the background noise and effectively filter out the ＂music noise＂, thus resulting in fairly good speech intelligibility and clarity.