引入幅度谱裁剪的语音增强算法

Speech Enhancement Algorithm by Introducing Amplitude Cutting

基于短时傅里叶变换的各种语音增强算法兼具效果良好及计算复杂度低的优点,得到广泛应用。这类算法利用功率谱进行计算,功率谱由幅度谱直接平方获得。本文提出一种通过裁剪幅度谱,以修正功率谱,提高语音存在处相对于噪声处的信噪比比值(信噪比对比度),从而恢复低信噪比语音的思路。将通过裁剪算法修正得到的功率谱用于基于短时傅里叶变换的语音增强算法,对于信噪比较低的语音位置,可以得到更好的增强效果。裁剪算法通过对小于一定阈值的幅度谱进行一定程度的衰减,再重新计算功率谱,使得能量高的位置的信噪比与能量低的位置信噪比的比值(信噪比对比度)得到提高,也就提高了能量高位置与能量低位置的区分度,有利于后续算法更准确地将高能量位置更准确地恢复出来。在时频域中,高能量位置通常代表着语音存在位置。添加幅度谱裁剪算法,能量高的位置被突出,也可以说是语音存在处被突出,故而可以获得更好的增强效果。文章最后给出了实验对比结果,语谱图及PESQ得分的对比结果,显示裁剪算法是有效的。

Speech enhancement algorithms based on short-time Fourier transform with the advantages of good effect and low calculation complexity, is widely used. The main parameter of this kind of algorithms is the power spectrum, which is the power of amplitude spectrum. In this paper, an ideal for cutting amplitude spectrum to correct power spectrum is presen- ted, which is supposed to improve the relative signal-to-noise ratio between voice existence and voice inexistence, and final- ly to restore the low SNR speech. Speech enhancement algorithms based on short-time Fourier transform would get a better result by introducing cutting algorithm to correct power spectrum. The ideal is to subtract a small value from the amplitude spectrum which below the threshold value, and to recalculate the power spectrum using the cut amplitude spectrum. Then the relative SNR between the high power position and the low power position could be improved. For the high power position usually stays for the speech existence, then the speech existence could be distinguished more easily from the speech inexist- ence. This benefits the following speech enhancement. The experiment results is given at the last of the paper. The spectro- gram and PESQ test show the presented algorithm worked.