Underdetermined blind speech separation of sparseness
基于稀疏性的欠定语音盲分离方法研究

A new sparseness-based method was proposed for mixing matrix estimation, in the case of poor sparseness of speech signals with increasing number of sources. The time-frequency bins with only one source were detected by Principal Component Analysis (PCA), and then were exploited to estimate the mixing matrix to improve the estimation performance. The proposed method is especially applicable to underdetermined blind speech separation. The reasons deteriorating the performance of blind speech separation were also pointed out. The simulation results demonstrate the conclusions above.