基于贪婪稀疏优化的欠定语音盲分离方法

Underdetermined blind speech source separation method based on greedy sparse optimization

提出一种基于语音信号稀疏特征的稀疏分量分析两步法,力图提高欠定情况下的语音信号盲分离性能。不同于传统的两步法,所提方法需要获取语音信号在变换域中的稀疏特征,将贪婪最优化思想引入至稀疏分量分析方法中,重构欠定盲分离语音源信号。通过仿真对比实验,展示了该方法应用于平稳声音信号和非平稳语音信号的盲分离效果,它能较好恢复语音源信号。与现有的最短路径法相比,所提算法可以提高两路以上观测信号的分离性能。相较于平滑L0范数算法,所提算法可以有效提高来波方向较近的语音盲信号分离性能。该方法具有更广阔的适用范围。

A two-step sparse component analysis method based on sparse feature of speech signal was proposed to improve the performance of underdetermined blind speech source separation.Different from the traditional two-step method,the proposed method needed to acquire the sparse features of speech signals in the transform domain,and greedy optimization idea was introduced into sparse component analysis method to reconstruct the underdetermined blind separation of speech source signals.Through the simulation and comparison experiments,the performance of the proposed method is shown by applying to stationary and non-stationary speech signals,and it can recover the speech source signals well.Compared with the existing shortest path method,the proposed algorithm can improve the separation performance of more than two observed signals.Compared with the smooth L0 norm algorithm,the proposed algorithm can effectively improve the separation performance of the unknown sources from the close-spaced direction.Therefore,the proposed algorithm has a broader scope of application.