A Modified Speech Enhancement Algorithm Based on the Subspace

A modified speech enhancement algorithm based on the subspace is advanced. It reduces residual noise caused by wrongly estimating noise eigenvalue matrix and speech eigenvalue matrix, because in the traditional speech enhancement algorithm based on the subspace, the eigenvalue matrix of noise are attained by eigen-decomposing to covariance matrix of noise, but covariance matrix of noise is estimated by using variance in the silence sequent, it cannot instead the whole noise, and lead to residual noise. Besides speech eigenvalue matrix equals the noisy speech´s minus noise´s, it increases residual noise. But the new approach tracks real time noise eigenvalue directly utilizing modified statistical information method, corrects speech eigenvalue matrix considering winner filter´s inhibitive ability to noise, then achieves the right speech eigenvalue matrix and restores a fortiori pure speech. The simulation results show that the modified algorithm advances signal noise ratio obviously and reduces residual noise evidently comparing with traditional method in the background of white noise.