基于EEMD和共振峰的自适应语音去噪  

Adaptive speech denoising based on EEMD and resonance peak

作　　者：李文志 屈晓旭[1] LI Wenzhi;QU Xiaoxu(College of Electronic Engineering,Naval University of Engineering,Wuhan 430000,China)

机构地区：[1]海军工程大学电子工程学院,湖北武汉430000

出　　处：《现代电子技术》2021年第23期52-56,共5页Modern Electronics Technique

基　　金：国家自然科学基金资助项目(61771483)。

摘　　要：针对传统的语音去噪方法可能滤除高频信息并且信噪比有进一步提升的问题,提出一种基于集合经验模态分解(EEMD)和共振峰的自适应语音去噪方法。语音信号通过EEMD分解成频率由高到低排列的本征模态函数(IMF)和余项,通过计算每个IMF分量和语音信号的第一共振峰对应频率值的距离,判断IMF的高低频部分,对高频信号进行小波阈值去噪后与低频信号相加得到重构信号;由于EEMD分解中添加的多组高斯白噪声在处理过程中不能完全中和,对重构后的语音信号的静音区产生较大影响,最后利用端点检测技术对静音区的残留噪声进行抑制。通过Matlab仿真结果表明,采用该去噪方法可以有效提高输出信噪比,并取得了良好的去噪效果。In view of the fact that the traditional speech denoising methods may filter out high-frequency information and further increase the signal-to-noise ratio(SNR),an adaptive speech denoising method based on the ensemble empirical mode decomposition(EEMD)and resonance peak is presented.The speech signal is decomposed by EEMD into intrinsic mode function(IMF)and remainder whose frequencies are arranged from high to low.By calculating the distance between each IMF component and the corresponding frequency value of the first resonance peak of the speech signal,the high and low frequency parts of the IMF are judged.After denoising by the wavelet threshold,the high frequency signal is added to the low frequency signal to get the reconstructed signal.The groups of Gaussian white noise added in EEMD decomposition can′t be completely neutralized during processing,which has a great influence on the mute area of the reconstructed speech signal.Finally,the endpoint detection technology is used to suppress the residual noise in the mute area.The results of Matlab simulation show that the method can effectively increase the output signal-to-noise ratio(SNR)and obtain good noise removal effect.

关 键 词：语音去噪 集合经验模态分解 共振峰 本征模态函数 信噪比 小波变换 小波阈值去噪 端点检测 

分 类 号：TN912.35-34[电子电信—通信与信息系统]