Abstract: An optimal speech denoising method which combines wavelet shrinkage and separation of unvoiced sounds and voiced sounds is presented.Firstly，the unvoiced sounds in speech often contain many high-frequency noise-similar components.If wavelet shrinkage method is directly adopted，it is likely to remove mistakenly these high-frequency components and result in distortion of speech，so it is necessary to separate unvoiced sounds and voiced sounds.Secondly，the optimal method of wavelet shrinkage is followed by optimization of many aspects such as different wavelets，threshold selection rules and threshold functions.Compared with the traditional method，it either removes noise as much as possible or retains the original characteristics of speech，which improves speech quality greatly.

Key words: wavelet denoising, threshold function, separation of unvoiced sounds and voiced sounds

摘要： 结合小波阈值去噪和清浊音分离技术，提出了一种优化的语音去噪新方法。首先，针对语音清音部分往往包含有许多类似噪声的高频成分的特点，对其直接进行小波阈值去噪很可能误除了这些高频成分，造成失真，因此有必要先对语音进行清浊音分离。其次，通过对不同小波函数、阈值选取规则以及阈值处理函数的优化，选择最佳的小波去噪方法。仿真结果表明，与经典小波阈值去噪方法相比，提出的方法既尽可能地去除噪声，又保留了原来语音的特征，较大地提高了语音质量。

关键词: 小波去噪, 阈值函数, 清浊音分离