In principle, you can use _any_ audio-file for convolution, so you _could_ use the plain samples (and maybe get some interesting results). For true impulse responses, you would have to treat the instrument like a room and record a sweep through its physical body.
I usually generate the test sweep and save it to a wav file. Then play the wav file with a wide-range speaker positioned at the source point I'm after while recording with a pair of omnis positioned at the perspective distance I'm after. The recorded wav file is then processed by the above tool into an impulse response file.
The above tool can also be used with any wav as an IR as Dietz observed, and you might come across an interesting effect. It's more common to do an IR of a piano, for instance, with a brick holding the sustain pedal down, and then using that IR to simulate the pedal-down effect with only pedal-up samples. The idea is to save having to sample each note twice, once with pedal-up and again with pedal-down.
I have done the oposite, however. Used a soundboard IR to lightly process a voice. It kind of simulates the effect I sometimes get when recording a singing piano player live.
1. Is there a better note to apply the convolution to? e.g is "C#" with all tone holes open or low "Bb", with all tone holes closed? (BTW I am trying to do this with a set of sax samples).
2. Should I try it with multiple notes and then use some process to get the average of the IRs?
3. Will I get better results if I use something like PRAAT to filter out the frequencies?
I know these are big questions and appreciate your help