S N ( ) =
1
2
X N ( )
N
1 I
Si ( )
I i =1
where e2 is the variance of the input white noise, and H ( ) is the filter frequency
response.
The filter coefficients and the variance of the input noise (referred to as the model
parameters) are evaluated by a linear predictive analysis. The spectrum estimates
based on such models tend to be more reliable than the classical methods, especially
when the length of the available data record is short.
To examine this in practice, perform LPC autocorrelation on 100 samples of the
random signal generated at the output of the recursive filter in section 2.1 (set the
LPC order to 2).
Take the 28-point FFT of the resultant analysis filter coefficients.
Display the produced spectrum in inverse (log = "-") and compare it with the true
power spectrum.
How does this estimate compare with that obtained using the Welch method (in
Section 2.1)?
Aladdin Ariyaeeinia
3