Just stirring you along a bit. 'Mace' and 'Mack' 'Macca' are just about all the McEdge cases.

If you look at the algorithm you will see it is just this:

      ($f) = /^(.)/;
      tr/AEHIOUWYBFPVCGJKQSXZDTLMNR/00000000111122222222334556/;
      ($fc) = /^(.)/;
      s/^$fc+//;
      tr///cs;
      tr/0//d;
      $_ = $f . $_ . '000';
      s/^(.{4}).*/$1/;
[download]

While there is no doubt it gives useful results it is exceptionally simplistic. The problem you are trying to solve is threefold:

1. Simple mis-spelling
2. Pronuncitation Differences
3. Dropped sylables

To a large extent the Knuth algorithm deals reasonably successfully with 1 and 2. Where it falls over is when there is either a mis-spelling of the first letter or a dropped sylable. As you could see in the example I fuzzed the matches by dropping symbols from the code to deal with this, given that there are only 6 numeric codes in use my estimate that it would on average pull 1% of the DB should have read more like 3% if you shorten the codes.

In my picture of a 'better' algorithm I imagine a mapping to perhaps [A-Z]{4} so you could perhaps do linear displacement rather than just dropping.....Oh well back to the real work Bayes filters are the soup de jour.