Removing parts of an audio clip containing silence (Strip Silence)

you need to incorporate 'hysteresis' into your detection:

once it falls below threshold, wait for a certain time before it can set a marker as the 'startSilence' marker(the timing is up to you, but you could start with trying something like 20ms of silence needed minimum). if the threshold goes back up before this time is up, then it's not truly a silence you want but just an abrupt change towards being below threshold.

if it satisfies the hysteresis condition, then you can mark that as a 'StartOfSilence' marker, then once it goes back above threshold after that, you can mark that index as a 'EndOfSilence' marker('Start...' and 'End...' can be variables in gen codebox, or stored within coll, etc.)... you don't need hysteresis for the 'End' since it should be following a significant amount of silence.

following that, you need to fade-out slightly before 'StartOfSilence' index, and fade-back-in slightly after 'EndOfSilence' index, and then you can cut the audio in between the fades by copying the audio starting at the fade-back-in point, all the way to the end of the buffer~, and paste that back right at the 'StartOfSilence' index.

then... rinse and repeat, rinse and repeat... until all silences are copied over :)

the process i described above is something i did within a Max-based tool i created, while working 2 years as full-time audio-ad producer at Pandora, because they had me on a very basic job cutting slates(the ID that comes before the ad, which is also followed by silence to be cut...) from client-ads, in between producing in-house ads. it was a shit job where i got badly exploited by the company, same as every other corporate american crap audio job i've ever worked :(

therefore, i don't like to think too hard about this type of thing, and i definitely won't remember what i did, but i'll upload the tool i made and perhaps looking at the C file here can help with a more advanced form(i believe i used RMS to detect things more reliably with hysteresis... but i don't want to explain the code... if it helps it helps, otherwise, ignore it)

___________________

there are probably much more efficient variations on what i described above(as in the way you detect, store variable/marker of index, and copy/cut... perhaps, you don't want to fade after End-Of-Silence if that might adjust a transient/attack to a kick-drum,etc... in which case, you can fade-in sooner and record a slightly different kind of index for the copy/cut-point(s)... but the general gist is the same: utilize hysteresis to detect the start, fade-out, detect the end, fade-back-in, cut/copy)... might even be more efficient in terms of speed to mark all silences and then do the multiple cuts/copies all in one go by utilizing an extra buffer... the extra buffer can help with resizing for the new total length, at the end, too.

i'm unable to think clearly on this(it's a depressing memory for me :( ...funny, how an algorithm can carry a feeling after you've been exploited to produce it)...but hopefully the general idea helps, and others might have better solutions or perhaps [mxj buf.Op] patches or similar that could help.

__________________

F--K Pandora radio! 😤

https://www.billboard.com/business/publishing/pandora-sued-mlc-unpaid-songwriter-royalties-late-fees-1235606144/

https://www.hollywoodreporter.com/business/business-news/comedians-pandora-licensing-fee-lawsuit-1235368741/

[Edit: @Cycling74, feel free to edit my post or ban me if the above goes against rules 😋🔫]