The following pure Perl code runs at least 2000 times faster than required to just keep up with one 1200 baud midi channel in one direction. Possibly fast enough?

use strict;
use warnings;
use Benchmark qw(cmpthese);

my @data28 = map {int (rand () * 256)} 1..28;

printf "%02x ", $_ for @data28;
print "\n";
my @xlate = thereAndBack (@data28);
printf "%02x ", $_ for @xlate;
print "\n\n";

my @data280 = map {int (rand () * 256)} 1..280;
my @data2800 = map {int (rand () * 256)} 1..2800;

cmpthese (-1,
    {
        data28 => sub {thereAndBack (@data28)},
        data280 => sub {thereAndBack (@data280)},
        data2800 => sub {thereAndBack (@data2800)},
    }
);


sub thereAndBack {
    return fromKarma (toKarma (@_));
}


sub toKarma {
    die "Blocks must be multiples of 8 bytes" if @_ % 7;
    my @raw = @_;
    my @karma;
    
    while (@raw) {
        my @block = splice @raw, 0, 7;
        my $extra = 0;
        
        for my $byte (@block) {
            $extra |= $byte & 0x80;
            $byte &= 0x7F;
            $extra >>= 1;
        }
        
        push @karma, ($extra, @block);
    }
    
    return @karma;
}


sub fromKarma {
    die "Blocks must be multiples of 8 bytes" if @_ % 8;
    my @karma = @_;
    my @raw;
    
    while (@karma) {
        my @block = splice @karma, 0, 8;
        my $extra = shift @block;
        
        for my $byte (reverse @block) {
            $extra <<= 1;
            $byte |= $extra & 0x80;
        }
        
        push @raw, @block;
    }
    
    return @raw;
}
[download]

Prints:

37 8b a0 23 ef 46 68 f8 a1 75 71 de aa eb 0e ee 08 54 c7 77 9f 0b ee d
+7 d3 f4 34 69 
37 8b a0 23 ef 46 68 f8 a1 75 71 de aa eb 0e ee 08 54 c7 77 9f 0b ee d
+7 d3 f4 34 69 

           Rate data2800  data280   data28
data2800  124/s       --     -90%     -98%
data280  1228/s     888%       --     -85%
data28   8200/s    6496%     568%       --
[download]

Note that there is some per packet overhead so there is some advantage in translating larger blocks.