为了刻画语音信号帧间相关性和使用更少的语音基表示语音特征,提出一种采用L_(1/2)稀疏约束的卷积非负矩阵分解方法进行单通道语音增强。首先,进行噪声学习得到噪声基;然后,以噪声基为先验信息结合L_(1/2)稀疏约束卷积非负矩阵分解方法学习含噪语音中的语音基成分;最后,利用学习到的语音基和系数重建出干净语音信号。在不同噪声环境下进行的实验结果表明,本文方法优于采用L_1稀疏约束的卷积非负矩阵方法及传统的统计语音增强方法。 In order to characterize the inter-frame correlation of speech signals and use fewer speech bases to represent speech features, a convolution non-negative matrix factorization method using L_ (1/2) sparse constraints is proposed for single-channel speech enhancement. Firstly, the noise base is obtained by noise learning. Then, the base components of speech in noisy speech are studied by prior knowledge of noise base and L_ (1/2) sparse constraint convolution nonnegative matrix factorization method. Finally, Voice base and coefficients reconstruct a clean speech signal. Experimental results under different noise environments show that the proposed method is superior to the convolutional nonnegative matrix method using L_1 sparse constraints and the traditional statistical speech enhancement method.