Computing SNRo from independent component analysis

[Ahm1979]

I work in speech denoising .I mixed clean speech (x) with AWGN .I used independent component analysis to separate the noise from the speech . I compute SNRi=10log10 var(x)/var. I compute SNRo=10log10 var(x)/var(s-x).s is the estimated of the clean speech after using independent component analysis.I get the same value of SNRo when SNRi changes from -10 to 15 dB . I think there is error in computing SNRo .I make normalized to the clean speech (x) because the amplitude of s is unity and amplitude of x is not unity but I do not get any improvement in SNRo .I need to explain how to compute SNRo after ICA in detail

 

[Audioguru]

How can you separate noise from speech when they have the same frequencies??
Maybe you use a highpass filter to cut lows then adult men sound like chipmunks. Then you use a lowpass filter to cut all high audio frequencies removing all the important consonant and sibilant sounds of speech making the speech difficult to understand?