1 Get a window of data: s[i], i=0,w-1 (w is the windowsize)
2 apply a high pass filter to the data (clicks contain a lot of high frequency amplitudes)
3 Compute hpf_mean = mean of hpf vector.
4 Moving backwards (j=i, i-50) through the windowed data, compute the first derivative dy[j] of the change in hpf from hpf[j+1] to hpf[j-1]
5 Compute dy_mean = mean(dy), dy_stddev = stddev(dy)
6 If dy[j] > 2.00*dy_stddev / sensitivity+dy_mean && hpf[j] > 2.0*mean_hpf / sensitivity then it's the start (end, that is) of a click.
7 If hpf[j] < 2.0*mean_hpf / sensitivity then it's the end (start, that is) of a click.
sensitivity is a factor that is set to make the detector more or less sensitive to click detection.The gymnasitics of looking at the means and standard deviations of the region around the sample of interest is to make the click detector smart and not get fooled by normal high frequency audio (like a cymbal crash), which is fairly long in duration, in essence, we're looking for outliers in the change of high frequency content to detect the end of a click, and looking for the high frequency amplitude to drop dramatically to detect the end of a click.