I recently tried csound's LPC stuff and I noticed a strange behaviour of 
of the lpanal utiliy. It seems
that the LPC analysis algorithm is having problems with absolut digital 
silence (all samples of
a analysis frame = 0). Lpanal then states
"Row 0 or 34 have maximum of 0
gauss: ill-conditioned"
and stops the analysis.
Having a quick look into the source code (lpanal.c) the gauss 
elimination algorithm first
searches the largest pivot and if it's below 1e-20 bails out.
I'm not quite sure about the math behind this, but if all of a frame's 
samples are 0, shouldn't
the filter coefficients and the linear prediction simply be 0, too?

cheers,
Jan Baumgart