Abstract: In order to overcome the problem that nonnegative matrix sparse decomposition is not enough, this paper puts forward a speech enhancement algorithm, which based on the nonsmooth nonnegative matrix factorization, by introducing smoothing matrix adjustment dictionaries matrix and coefficient matrix sparsity. Firstly it constructs combined dictionary matrix through speech and noise prior dictionary learning, and then the speech is enhanced by the use of nonsmooth nonnegative matrix decomposition to update the combined dictionary matrix projection coefficient of the noisy speech , and update the prior noise dictionary by sliding window. Simulation results show that the proposed algorithm has better ability to suppress the noise than the nonnegative matrix factorization speech enhancement algorithm and MMSE algorithm.

Key words: nonsmooth nonnegative matrix factorization, sparsity, speech enhancement

摘要： 针对非负矩阵分解稀疏性不够，通过引入平滑矩阵调节字典矩阵和系数矩阵的稀疏性，提出基于非平滑非负矩阵分解语音增强算法。算法通过语音和噪声的先验字典学习构造联合字典矩阵；然后通过非平滑非负矩阵分解更新带噪语音在联合字典矩阵下的投影系数实现语音增强；同时通过滑动窗口法实时更新先验噪声字典。仿真结果表明，该算法相对非负矩阵分解语音增强算法和MMSE算法具有更好的抑制噪声能力。

关键词: 非平滑非负矩阵分解, 稀疏性, 语音增强