论文部分内容阅读
文章提出了一种基于非负矩阵分解的语音增强算法。该算法包括两个阶段,训练阶段和增强阶段。训练阶段通过非负矩阵分解算法对纯净的噪声频谱进行训练,得到噪声字典矩阵,保存其作为增强阶段的先验信息。增强阶段首先通过非负矩阵分解算法对带噪语音的频谱进行分解,然后联合噪声字典矩阵和推导得到的相应迭代公式对语音字典矩阵和语音编码矩阵进行估计,重构增强语音。仿真结果表明,文中增强方案在抑制背景噪声,提高信噪比和减少语音失真方面要优于传统的语音增强算法。
This paper presents a speech enhancement algorithm based on nonnegative matrix factorization. The algorithm includes two phases, a training phase and an enhancement phase. In the training phase, the pure noise spectrum is trained by the nonnegative matrix factorization algorithm, and the noise dictionary matrix is obtained, and the priori information is saved as the enhancement phase. In the enhancement phase, the spectrum of noisy speech is first decomposed by non-negative matrix factorization algorithm, then the speech dictionary matrix and speech coding matrix are estimated by combining the noise dictionary matrix and the derived corresponding iterative formula to reconstruct the enhanced speech. The simulation results show that the proposed enhancement scheme is superior to the traditional speech enhancement algorithm in suppressing background noise, improving signal-to-noise ratio and reducing speech distortion.