You made the mistake of modifying $string directly so that in all but the first call, there are NO characters that need to be transliterated so it is much faster. Once that is fixed, it doesn't have such a big lead.

Whoops! You're right, I hadn't expected it to modify $string in-place. I suppose that's due to Benchmark imposing a void context on the return.

My lesson learned today: Never trust your own benchmarks :)