But.... It doesn't look like it needs to be realtime.

I'd think about playing the file once and storing the sample locations where the volume drops below a treshold (as well as the ones where it comes back) in a array, then playing it a second time and skipping those sections while writing the results to a new file. in "silent" mode this should be doable faster then realtime.

You could also make a array of the length of the file (in samples) and only write buf.last() => array[sample++] as long as it's over the treshold. Once the buffer reaches it's end you'd playback the array again and record that, taking care to stop after the highest value that "sample" got to in the first pass. This would be CPU intesive but it's quite clean as a method, I think.

It all stands or falls with the quality of the envelope follower.

I've been thinking about doing something similar to the first option and pre-scanning wavefiles for positive zero-crossings and storing those in a array. This would take extra time when loading a file but from then on you could do very clean loops, for example to loop just a few cycles. I wonder how samplers that have a "snap to zero crossings" setting do this. It might still be quite tricky because not all samplers that do this do it all that well....

Hope that's of some use.
Kas.