filtering in general may not be the best approach because some of your partials from one xylophone note will overlap with other note's partials. They are inharmonic complex tones, which are not so easy to predict.... you'll probably have to measure the frequencies of each note of your xylophone to know exactly what the spectrum is like.
In terms of averaging like Jamie suggested... suppose you want to compute the expectation of the power spectral density. You would take the fft of the auto-covariance of your recieved signal, divided by the number of blocks in your time frame. (dividing by a number of blocks will not in general be necessary, when all you need to do is find a peak, with pique~ as before) This will give you a very clear/accurate peak, without much jitter/noise to clean up.
Chuck