Recommended AR Processing Chain

Based on the above tests, we now provide the recommened chains for AR/ACF models. Let $ {\bf x}$ be an $ N\times M$ matrix of data segments ($ M$ segments of length $ N$). The recommended processing chain for AR features using circular spectral model is the following:
   E0=0; % indicate that we use r0, not sig2
   jout = 0;
   [a,jout]=module_ar_mlx(x,jout,P,E0);
   [k,jout]=module_ar2rc(a,jout);
   [z,jout]=module_bilinear(k,jout);
This can be inverted using:
   E0=0; % indicate that we use r0, not sig2
   k = module_bilinear_synth(z);
   r=module_acf2rc_synth(k,E0);
   y=module_acf_synth(r,N);
   x=module_dftmsq_synth(y,N);
Th use $ \sigma^2$ instead of $ r_0$, remove the call to module_as2r0 in the feature chain If intermediate spectrum is needed, use the alternative chain:
   E0=0; % indicate that we use r0, not sig2
   [y,jout]=module_dftmsq(x,jout);
   [a,jout]=module_ar_ml(y,jout,P,E0);
   [K,jout]=module_ar2rc(a,jout);
   [z,jout]=module_bilinear(K,jout);
although it is a tiny bit slower. For non-circular spectral models, use
   [r,jout]=module_acfx(x,jout,P);
   [k, jout] = module_acf2rc(r, jout);
   [z, jout] = module_bilinear(k, jout);



Baggenstoss 2017-05-19