Abstract: The basic spectral subtraction algorithm has some shortcomings under the low SNR environment, such as, poor noise reducing effect, too much residual musical noise and low speech intelligibility. An improved spectral subtraction algorithm is presented. At first, the cepstral distances of speech signal is calculated to distinguish the voice from the noise part. In this process, a dynamical calculation value of noise is adopted in the improved algorithm while the basic algorithm uses a statistical mean value. Then according to the ratio of cepstral distance of current frame to noise frame, it can set a dynamical spectral values coefficient, which avoids the shortcomings caused by the unchanged coefficient in basic algorithm. Finally, three methods are proposed to reduce the musical noise. The numerical results show that the improved algorithm can reduce noise effectively, and also enhance speech intelligibility, which can meet the demand of speech enhancement.

Key words: spectral subtraction, cepstral distance, speech enhancement, musical noise

摘要： 针对低信噪比条件下基本谱减算法存在降噪效果不佳，产生音乐噪声过大，语音可懂度不高的问题，提出了一种改进型的谱减算法。算法先计算语音信号的倒谱距离值，检测出噪音段和语音段，用动态计算的噪声值代替基本谱减法采用的噪声统计均值；根据当前帧和噪声帧的倒谱距离比值动态设置谱减系数，改进了传统算法中谱减系数保持不变的缺点；同时采用三种方法抑制音乐噪声。仿真实验表明，在低信噪比情况下，改进型的谱减算法可以有效降噪，提高信噪比和可懂度，达到语音增强的目的。

关键词: 谱减法, 倒谱距离, 语音增强, 音乐噪声