This may not be an answer per se, but as a software developer and musician I'd like to shed some light on why I think you would not be satisfied with any software claiming to do this.
WHY IT PROBABLY IS NOT GOING TO WORK AS GOOD AS YOU NEED IT TO:
In order for the software to decide what midi note to play,
it will need to analyze enough data in order to make a good choice.
This will introduce a severe latency.
A POSSIBLE SOLUTION:
I think, the best solution for anything like this would be to prerecord what you think a HH sounds like, a bass drum, a ride, crash and snare. If you could keep the drum count low, and assure that each sound is significantly different from all the others then it could probably be done with less than 10ms latency.
You'll need fewer samples differentiating between a bass drum and a high hat, than the tim, and the tam.
If the software required you to enter some samples of yourself beat-boxing each drum sound; the problem domain would no longer be that of matching your sound to a best guess over a mathematical and statistical representation of a 'snare drum sound', but rather a much simpler approximation, trying to find a close match between your sound and a prerecording of you for that same sound.
There is also the problem of sample hold. A good heuristic for this might be to simply wait to send the MIDI OFF message until there is silence on the mic. That way, you could hold a bass drum for as long as you had breath.
There might be some programs out there that can analyze an audio stream
for certain sounds that you enter yourself, and trigger a reaction to those.
If that reaction could be to send a MIDI message you've got a solution.
Actually, as I wrote this answer it seems like an interesting and challenging project. Let me know if you still have not found anything in a few days time, will you?
