Abstract: Speech enhancement aims at extracting clean speech from noisy speech. Traditional speech enhancement algorithms proposed for continuous noise can not work well when clean speech is polluted by impulsive noise. The feature of the distribution of impulsive noise in time domain has to be concerned for speech enhancement. This paper provided a new speech enhancement method for the removal of noise when clean speech is polluted by stationary impulsive noise. By analyzing the samples of impulsive noise and noisy speech, the method found the frequency band in which the energy of impulsive noise was most evident compared with the energy of noisy speech. Furthermore, filter banks were used in recognizing whether the speech frame was polluted. The method based on Kalman filtering removed impulsive noise from the polluted signal frame. Compared with the traditional algorithms based on Kalman filtering, the method also improved the process of the parameters estimation of Auto Regressive (AR) model. In some speech enhancement tests, speech was polluted by white impulsive noise or colored impulsive noise, the results show that the method has achieved an improved Signal-to-Noise Ratio (SNR) and removes impulsive noise even when the SNR of noisy speech is low.

Key words: impulse noise, speech enhancement, filter bank

摘要： 语音增强的目标在于从含噪信号中提取纯净语音，纯净语音在某些环境下会被脉冲噪声所污染，但脉冲噪声的时域分布特征却给语音增强带来困难，使传统方法在脉冲噪声环境下难以取得满意效果。为在平稳脉冲噪声环境下进行语音增强，提出了一种新方法。该方法通过计算确定脉冲噪声样本的能量与含噪信号样本的能量之比最大的频段，利用该频段能量分布情况逐帧判别语音信号是否被脉冲噪声所污染。进一步地，该方法只在被脉冲噪声污染的帧应用卡尔曼滤波算法去噪，并改进了传统算法执行时的自回归(AR)模型参数估计过程。实验中，采用白色脉冲噪声以及有色脉冲噪声污染语音信号，并对低输入信噪比的信号进行语音增强，结果表明所提出的算法能显著地改善信噪比和抑制脉冲噪声。

关键词: 脉冲噪声, 语音增强, 卡尔曼滤波, AR模型, 滤波器组