While maybe not technically correct, we actually only care about the difference
between onsets (any base constant is removed). This should work better since it
determines when the IR is most audible, whereas previously it used a variable
threshold of when it reached 15% of the max amplitude.
An even better method may be to check where the IR amplitude exceeds a fixed
threshold (i.e. the same threshold applied to all IRs), but that would need
tweaking to find a level that doesn't catch random noise and doesn't
potentially miss the more occluded IRs.
f63880c2ff