Robustness Techniques for Feature Extraction - Berlin Chen
Robustness Techniques for Feature Extraction - Berlin Chen
Robustness Techniques for Feature Extraction - Berlin Chen
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Spectral Subtraction (SS) S. F. Boll, 1979<br />
• A Speech Enhancement Technique<br />
• Estimate the magnitude (or the power) of clean speech by<br />
explicitly subtracting the noise magnitude (or the power)<br />
spectrum from the noisy magnitude (or power) spectrum<br />
• Basic Assumption of Spectral Subtraction<br />
– The clean speech s [ m]<br />
is corrupted by additive noise n[ m]<br />
– Different frequencies are uncorrelated from each other<br />
– s [ m]<br />
and n[ m]<br />
are statistically independent, so that the power<br />
spectrum of the noisy speech x[ m]<br />
can be expressed as:<br />
P ( ω ) = P ( ω ) + P ( ω )<br />
X<br />
S<br />
N<br />
– To eliminate the additive noise: P ( ω ) = P ( ω ) − P ( ω )<br />
S<br />
X<br />
N<br />
– We can obtain an estimate of P N<br />
( ω ) using the average period of M<br />
frames that known to be just noise:<br />
Pˆ<br />
1<br />
M<br />
M<br />
( ) ∑ − 1<br />
ω = ( ω )<br />
P N N ,i i=<br />
0<br />
14