A quick trial of 16K19: I just put together my best
two rate 1/8 K=19 codes to make a 1/16 K=19 polynomial
set and ran a comparison test using 12 char messages
at -0.5dB Eb/N0.

Decode success rates with optimum list length:

 8K19A    69.8 %
 16K19T1  67.4 %
 8K21A    74.1 %

This is typical.  The two 1/8 codes are not dissimilar
enough to see a performance gain when used together.
It usually takes many trials of different combinations
to find one that shows some coding gain.  I have 147
reasonable 1/8 K=19 codes, so plenty of permutations
to try.   The method I use is to put pairs of 1/8
codes together and calculate the first 30 or 40 terms
of the distance spectrum.  The pairs that have the
best spectrum then get a full test.

I can also add to the list of 8K19 codes by combining
4K19 and so on.  This is a lot quicker than just testing
random 1/16 polynomials.

It's nearly 3 years since I last ran the 'code factory'
and I can probably make some improvements to the process
now.  Certainly the PCs are more powerful.  Incidentally
the 'A' in 8K19A just labels a particular polynomial.
If a better one turns up, it'll go in the menu as 8K19B
and so on.

--
Paul Nicholson
--