稀疏低秩模型下的单通道自学习语音增强算法

发布时间：2018-03-15 23:35

本文选题：语音增强　切入点：无监督学习　出处：《数据采集与处理》2014年02期 　论文类型：期刊论文

【摘要】：针对现有基于字典学习的增强算法依赖先验信息的问题,基于矩阵的稀疏低秩分解提出一种无监督的单通道语音增强算法。该算法首先通过稀疏低秩分解将带噪语音的幅度谱分解为低秩、稀疏和噪声3部分,然后通过对低秩部分进行自学习构建出噪声字典,最后利用所得噪声字典和乘性迭代准则于低秩和稀疏部分中分离出纯净语音。相较于其他基于字典学习的语音增强算法,本文所提算法无需语音或噪声的先验信息,因而更加方便和实用。实验结果显示,本文算法能够在保留语音谐波结构的同时有效抑制噪声,增强效果明显优于鲁棒主成分分析和多带谱减法。
[Abstract]:According to the existing dictionary learning enhancement algorithm Apriori information based on sparse matrix low rank decomposition of a single channel speech enhancement algorithm based on unsupervised algorithm. The first through the low rank sparse decomposing the noisy speech spectrum is decomposed into low rank, sparse and noise of the 3 part, based on low rank part of the self-learning construct noise dictionary, finally using the dictionary noise and multiplicative iterative criterion for low rank and sparse part separated pure speech. Compared with other speech enhancement based on dictionary learning algorithm, the proposed algorithm without speech or noise prior information, which is more convenient and practical results. This paper shows that the algorithm can effectively restrain the noise while retaining speech harmonic structure at the same time, enhance the analysis is better than robust PCA and multi band spectral subtraction.

【作者单位】：解放军理工大学指挥信息系统学院;
【基金】：江苏省自然科学基金(BK2012510)资助项目
【分类号】：TN912.3

【参考文献】