The issue I wrote on does not seem related to byte ordering,
because the auxiliary information - rms energies, pitch, error,
get read accurately. In order to make sense of the data, 
I've been doing analyses with as few as two poles. The
imaginary components are sometimes pi or its sub-multiples -
not something which you'd expect to see with z transforms.

The SDIF format sounds good. However, the support for it
seems not to have gained any momentum, in the couple of
years the page at CNMAT has been up. I hear that Xavier
Serra's SMS uses it. I was at the CNMAT site a few days
ago, and note that many aspects of the full specification
are still not done.