弱稀疏语音信号的欠定盲分离

Underdetermined blind separation for weak sparse speech signals

摘要:

为解决弱稀疏语音信号的欠定盲分离问题，根据语音信号的部分W-分离正交性，提出一种基于单源主导区间的混合矩阵盲估计方法。该方法根据单源主导区间的性质，通过二元行矢量提取单源观测样本，对单源观测样本进行K均值聚类和主成分分析来估计混合矩阵。仿真结果表明，提出的方法可有效提高分离语音的性能，与直接利用K-PCA方法相比，分离语音的平均信噪比提高了10 dB左右。

Abstract:

In order to solve the problem of underdetermined blind speech separation for weak sparse speech signals, according to the partial approximate W-disjoint orthogonality of weak sparse speech signals, a blind mixing matrix estimation method is proposed based on single-source dominated areas. In this method, according to the property of the single-source dominant areas, single-source observational samples are extracted by a binary row vector; then the mixing matrix is estimated by K means clustering and Principal Component Analysis(PCA) on extracted samples. Simulation results show that the proposed method can effectively improve the performance of separated speech signals, and its averaged Signal To Noise Ratio(SNR) is improved by 10 dB compared to K-PCA method.