Since speech is slow time-varying process, an accu-
rate set of predictor coefficients is adaptively deter-
mined over short intervals (10 ms ~30 ms) called
frames, during which time-invariance is assumed.
The autocorrelation method and the covariance
method are two standard methods of solving for the
predictor coefficients. [1, 2]
A robust solution technique will result in the vo-
cal-tract information being captured by Hz, wheth-
er speech is clean or corrupted by noise and/or
channel effects. Then, the predictor coefficients
would either be invariant or show very little vari-
ation when speech is corrupted. Subsequently, the
features would be naturally robust. [3]
Another attempt at representing the speech spec-
trum involves an approximation that gives more
emphasis to those frequencies that have greater au-
ditory prominence.
43