基于感知小波变换的语音增强方法研究

在ERB尺度下构造的感知小波符合人耳对固有语音的频率感知特性，通过一种纯数学算法计算其参数，在听觉感知上可以近乎完美地使信号进行重构。首先采用感知小波对带噪语音进行分解，其次在语音信号的子带层次上用一种类似于软阈值的无穷阶可导的函数进行阈值处理，最后应用谱减法进行二次增强。实验表明，该算法使信噪比和PESQ得分都有较大提高，特别是在信噪比较高时，语音具有很好的清晰度和可懂度。

Research on speech enhancement based on perceptual wavelet transform

A perceptual wavelet constructed under ERB scale accorded with the perceptual of human hearing system, used pure mathematics method to calculate its paraneter and the perceptual wavelet could reconstructe faultlessly in auditory perception. Used perceptual wavelet to decompose noisy speech. Used an infinite-order derivative function which was similar to the soft threshold on sub-band level of spech signals, then used spectral subtraction in the second time speech enhancement. Experiment result shows the algorithm is highly improved both SNR and PESQ scores, especially when the SNR is high, the speech has well definition and intelligibility.