CAJ | 학술논문

针对语音信号的弱时频正交性,提出一种基于主分量分析的混合矩阵估计方法.在时频域中,允许每个时频点存在任意多个源信号,通过对每个时频点进行主分量分析,检测只有一个源信号存在的时频点,此类时频点最大特征值对应的特征向量即为混合向量的一个估计,因此对所有估计出的混合向量进行K均值聚类,将聚类中心作为混合矩阵的估计.实验仿真表明,提出的方法提高了混合矩阵的估计精度,特别适用于估计欠定情况下的混合矩阵.
침대어음신호적약시빈정교성,제출일충기우주분량분석적혼합구진고계방법.재시빈역중,윤허매개시빈점존재임의다개원신호,통과대매개시빈점진행주분량분석,검측지유일개원신호존재적시빈점,차류시빈점최대특정치대응적특정향량즉위혼합향량적일개고계,인차대소유고계출적혼합향량진행K균치취류,장취류중심작위혼합구진적고계.실험방진표명,제출적방법제고료혼합구진적고계정도,특별괄용우고계흠정정황하적혼합구진.
In blind speech separation, a method based on principal component analysis (PCA) is proposed to estimate the mixing matrix for the weak time-frequency orthogonality property of speech. In the time-frequency domain, the proposed method allows the arbitrary number of sources to be existed in a time-frequency bin, then PCA is applied to every time-frequency bin to detect the existed one source in the time-frequency bins. In the detected time-frequency bins, the eigenvector associated with the maximum eigenvalue is an estimation of the mixing vectors, so K-means clustering is exploited on all the mixing vectors and the cluster centers are used as the estimation of the mixing matrix. Simulation results demonstrate that the proposed method can improve estimation precision, especially for estimating the mixing matrix in underdetermined case.