The signal passes so many buffers and code.
The resulting latency is hard to predict.
Did you know the Arduino can send midi codes itself ?
I think that most latency is in the firmata path. So I'm guessing that the latency could be easily half of what you have now by directly sending midi strings.
Detecting a knock is also something that often can be made a lot faster.
If you calculate the delta and test for a certain level, the knock will be detected on the rising slope. You can't get faster than that.