Rezayee's and Gazor's noise model [8]

$$y[n]=x[n]+b[n]$$

$$R_{x}=U \Lambda_x U^T$$

Ephraim and Van Trees [4] assume $R_{b} \approx \sigma^2 I$. Rezayee and Gazor assume that the noise and speech eigenvectors are approximately equal. Then

$$R_{b}=U \Lambda_{b} U^T$$

where $\Lambda_{b} \approx diag(\lambda_{b1}, \lambda_{b2}, \ldots, \lambda_{bK})$. This is especially true if $R_b$ is close to Toeplitz.

Noise samples taken from the most recent silent period are projected onto the current signal eigenvectors to obtain the variance of each noise component.