Interesting.
I've actually been working on this and I've emailed various experts
on the subject, but didn't exactly get any "one-liners" :)
Using: http://www.phon.ucl.ac.uk/resource/sfs.htm I used voc26
"voc26 is a straightforward implementation of a 26-channel
filterbank analyser using two second-order section Butter-
worth bandpass filters spaced as in Holmes paper in IEE
proceedings with SRU extensions at the low and high fre-
quencies. Energy smoothing done at 50Hz to give default
100Hz framerate"
Hopefully that means more to you than it does to me...
I then processed various cd samples through 8hz-mp3, lame & lamer
at 32, 56, 128 bits and vbr (for lamer), reduced the voc26 averages
to coarse samples and differenced the data with good results for
56 bits and above. I then went to www.napster.com and pulled an
elton john song that I already had from three different people
and it failed miserably. With sfs you can visualise the waveform
and just looking at the samples (first 5s or so) there were large
amplitude variations, one inverted waveform and one commenced
the song without the short silence that the others had!
To progress I needed to normalise the data to hopefully remove
the amplitude variations and determine a characteristic start
point. I hadn't got around to that yet....
Paul