That work has already been done. MIDI-to-sound latency is 3.5 to 7ms and it's not likely to get any better than that. Tsunami is a polyphonic player and requires an audio buffer to mix multiple voices - just like DAW software running on a PC. The currently latency represents the best balance between all the realtime tasks Tsunami has to manage and ensuring that there are no clicks or pops due to a starved buffer.
Hope this helps.