Given multiple audio files, can I extract background music that has different sound effects over it?

Question

Let's imagine I have a movie out of which I would like to extract a soundtrack (i.e., a piece of music). The problem being that the music is, of course, "tainted" with various sound effects and actors talking.

But now, suppose I have two or more clips of the same music, but from different scenes, meaning that the sample has different sound effects, but the same music in the background.

It should be possible to manually try and cut this music together (but the result would likely be horrifying). So I am wondering: Is there an automatic way / software / algorithm to extract the music from multiple such samples, whilst filtering out all the noise (i.e. spoken words, sound effects, etc.) in an acceptable quality?

Answer 1

If I understand, you're asking for an automated way to distinguish music from other sound after the two have been mixed together. There's a lot of tech magic out there these days, so I may not be aware of something that can do this, but in general what you want sounds unachievable.

But if you have multiple identical tracks with 'clear' spaces in different areas, I'd stack them up in multiple layers and gate them all off except the one that's currently in the clear. It is manual labor, but this should be fairly quick if the clear stretches are at least a few seconds long each.

Answer 2

If it is exactly the same, you should be able to cancel out the music, leaving both effects, but I'm not sure about extracting it. If you have A + B and A + C (where A is music and B and C are the sound effects), then when you invert one and subtract it, you get B-C (ie, the sound effects from one overlaping the sound effects of the other, but with a phase inversion.)

You still don't have any simple way to identify what makes up B vs C because all you have is A+B and A+C. There may be some complex analysis that could be done of B-C to figure out what could fit in to A + B and A+C and still be consistent with B-C, but I'm not sure how it would work or how well it would do.

Note, I'm just answering from a purely theoretical perspective, I have no direct experience in trying to do this, so there may well be an advanced analysis tool I don't know of. I just know the simple approaches don't help.

The best you can do with basic sound is to reduce the volume of the effects and increase the quantity. If you have, say, 4 different instances of the song with different effects, you can sum them as such:

A+B + A+C + A+D + A+E = 4A + B + C + D + E (4A + B + C + D + E)/4 = A + .25B + .25C + .25D + .25E

Note, this only works if you have pretty exactly or very closely similar copies of the music though. If you don't, they won't sum properly and will instead interfere and it won't work.

Answer 3

Just throwing this in as a quick third possibility - that the original 5.1 soundtrack might [if you're lucky] already have most of the speech & perhaps even some of the effects in the centre channel only, giving you 2 cleaner channels (front L & R) to try working on.

Failing that, of course, there's always the soundtrack album...