its been done, quite a few articles... though many not publically available, but this one gives an insight, and said it worked quite well doc link
that doesn't matter, the point of using something like a FF neural net is you do offline training, i.e. do the training on your desktop , then once 'complete' transfer the training data to the axoloti to execute. the normal execution is much lower in cpu consumption, as its basically just lots of multiplications ... though i don't know, even this low consumption, is still low enough for axoloti.
also note: creating training sets can be pretty time consuming, as you need many examples to make it work ... that said i wonder, if in this case you could automate the creation of the training sets?
e.g. play lots of notes in on your guitar, then auto slice it , and run it thru a slow but accurate pitch detector.
this could work quite well, as another thing about training is you need to randomise the order the notes are played, otherwise the NN will just start predicting what you are going to play next