Uh oh, you've hit a nerve
I can't help giving the full story, for those who may be interested. This voicing phenomenon (-tori -> -dori) actually has a name in Japanese: "rendaku" (the word Jun gave, "nigori", actually just means to voice an otherwise voiceless consonant). The rule that governs it is sometimes called Lyman's Law, because it was first written about in English by an American of that name, about 100 years ago.
The way it works is actually fairly simple, although most Japanese don't know it (consciously) and would have a hard time figuring it out if you asked them. In a compound word, an initial consonant in the second element of the compound will often be voiced as long as there is no other voiced obstruent in that second element. For those who read Japanese, the way you recognize a voiced obstruent is that it's got a nigori mark (two dots or small circle) when written in kana. Other voiced sounds such as m, n, r, w do not trigger the rule.
It is also more a "negative" type constraint than a rule that tells you exactly when to add voicing. In other words, if there is a voiced obstruent later in the word, rendaku is blocked, however if there is no such sound present, sometimes the initial consonant will be voiced and sometimes not. Thus the fact that both -tori and -dori are permissible pronunciations for attacks involving a hold. I believe that in these cases where rendaku is not blocked by the rule, whether to voice or not maybe sometimes be indicated lexically (i.e. it comes built-in to how the word is stored in native speakers' heads), or sometimes may be controlled by dialect variation.
Hope this has been interesting for the non-linguistics geeks among you