基于稀疏约束的PLCA和谱掩蔽的语音增强算法

针对传统语音增强方法在非平稳噪声环境下增强效果不理想的问题,提出了一种基于稀疏约束的概率潜分量分析(PLCA)和谱掩蔽的语音增强算法。该算法分为训练和增强两个阶段。训练阶段用稀疏约束的PLCA(SPLCA)和无约束的PLCA分别对语音谱和噪声谱建模成意义清晰的边缘分布,并用期望最大(EM)算法求其最优边缘分布,得到语音字典和噪声字典。增强阶段固定训练的字典,利用SPLCA推导出对应的语音编码矩阵和噪声编码矩阵,初步重构出语音和噪声,最后利用谱掩蔽得到增强语音。实验结果表明,该算法在抑制噪声、提高信噪比和减少语音失真方面要优于传统方法。

Single Channel Speech Enhancement Based on PLCA with Sparsity Constraints and Spectral Masks

As previous single - channel enhancing algorithm of speech is not ideal under the condition of non - stationary noise environment, a new algorithm based on probabilistic latent component analysis （PLCA） with sparsity constraints and spectral masks is proposed in this paper. This algorithm contains mainly two steps. In the first step, modeling speech spec- trum and noise spectrum respectively into marginal distribution of clear meaning by using PLCA with sparsity constraints （SPLCA） and PLCA. Meanwhile, optimal marginal distribution is sought by utilizing expectation -maximization algorithm （EM） so that deriving the dictionaries of speech and noise. In the second step, fixing dictionary trained is followed by taking use of SPLCA to infer corresponding coding matrix of speech and noise to reconstruct preliminarily the speech and noise, ac- quiring enhanced speech by spectral masking eventually. Experimental results show that the algorithm proposed in the paper performs better than the traditional speech enhancement algorithm in suppressing non - stationary noise, raising SNR and re- ducing speech distortion.