传统的子空间语音增强方法在推导过程中,一般假定噪声的方差矩阵是满秩的.在窄带噪声或者音调噪声下这一条件就不能满足,文中采用直接对噪声信号进行子空间白化的方法,将传统的子空间语音增强方法推广到了窄带噪声的情况,仿真结果表明了该算法的有效性。
A single channel speech enhancement method for noisy speech signals is proposed based on the subspace method. Conditional subspace based enhancement methods require that the noise covariance matrix is full-ranked. Unfortunately this is not satisfied in case of narrowband noise or tonal noise. The proposed method directly deals with rank-deficient noise using a subspace whitening method. Experimental results demonstrate the effectiveness of the proposed algorithms.