I was thinking of something along the lines of

$soundex{ $name } = soundex($1) 
    if $name =~ m[^(?:O'|Mac|Mc)([A-Z].*$)];
[download]

Which would only work if the list is nicely capitalised, which going by the few BazB posted, it seems that his data might be. The idea was that a search for Connor or Keefe, would also find O'Connor and O'Keefe, which using unassisted soundex wouldn't find.

#! perl -slw
use strict;
use Text::Soundex;


printf "%20s : %s : %s\n",
      $_, m[^(?:O'|Mac|Mc)([A-Z].*$)]  ? soundex($1) : soundex($_), so
+undex($_)
      for qw[
          Connor O'Connor
          Keefe O'Keefe
          MacDonald McDonald Donald Donaldson O'Donnell Donagal O'Dona
+gal
      ];

__END__
P:\test>test
              Connor : C560 : C560
            O'Connor : C560 : O256
            
               Keefe : K100 : K100
             O'Keefe : K100 : O210
             
           MacDonald : D543 : M235
            McDonald : D543 : M235
              Donald : D543 : D543
           Donaldson : D543 : D543
           O'Donnell : D540 : O354
             Donagal : D524 : D524
           O'Donagal : D524 : O352
[download]

The first column of soundex codes are the assisted ones, the second unassisted. You can see what a difference it makes.

That said. It would screw up sound alikes Magee and MacGee, so it may not be such a good idea. Another possibility would be to store two codes for some names, but then you move into needing to normalise the soundex codes into another table. Which wouldn't be a bad thing, but does add complication.