针对现有基于字典学习的增强算法需要先验信息、不易实时处理的问题,提出一种便于实时处理的无监督的单通道语音增强算法。首先,该算法将无监督条件下背景噪声的建模问题转化为带噪语音幅度谱的稀疏低秩噪声分解;然后,采用增量非负子空间方法对背景噪声进行在线字典学习,获得能够体现背景噪声时变特性的自适应噪声字典;最后,利用所得的噪声字典,采用易于实时处理的逐帧迭代方式,对带噪语音进行处理。实验结果表明:相较于多带谱减法和基于低秩稀疏矩阵分解的增强算法,所提算法在噪声抑制方面的性能尤为显著,在多项性能评价指标上,均表现出更好的结果。 Aiming at the problem that the existing algorithm based on dictionary learning needs prior information and is not easy to be dealt with in real time, an unsupervised single-channel speech enhancement algorithm is proposed for real-time processing. Firstly, the algorithm transforms the modeling problem of background noise under unsupervised conditions into the sparse low-rank noise decomposition of the noisy speech amplitude spectrum. Then, the algorithm uses incremental non-negative subspace method to conduct on-line dictionary learning of background noise, Background noise time-varying adaptive noise dictionary; Finally, the use of the resulting noise dictionary, using real-time frame-by-frame iteration, the noisy speech processing. The experimental results show that compared with the multi-spectral subtraction and the enhancement algorithm based on the low-rank sparse matrix decomposition, the performance of the proposed algorithm is particularly significant in noise suppression and shows better results in a number of performance evaluation indexes.