Getting Two Audio Files to Sync

[Lio]

Okay, here's a tricky one. I have two copies of the same music recording from Harry Potter 3. One copy has a different instrument mixed in that I'd like to isolate as much as possible. To do that I first need to line the two recordings up pretty much exactly. The mixing between the two is different so comparing the waveforms isn't really an option.
After playing with it the past couple of days in Audacity, I've gotten it very close to syncing (I lined it up via an audio glitch about 21 seconds in) but am just off by a little. I opened my result up in RX7 and made an interesting discovery:

You can see there's some sort of filter-sweep type thing. The shape of it confuses me. Does this mean that one of the files is running slightly faster than the other, and that the area 20 seconds in is where the two files are "meeting" before going out of sync again?

I can PM anyone the Audacity file to anyone who's interested, or I can post it here if that's allowed. Thanks for any insight.

[deleted user]

What exactly did you do, mix the two files? Is that sweep also in the source files or not? And what was your goal with this, out of curiosity?

Generally yeah, I can imagine something like this resulting from speed differences, leaving various frequencies to cancel each other out a little when their wavelength harmonizes with the amount of desync that changes over time. It starts in higher frequencies and goes down, so from shorter to longer wavelengths, which could correspond with a desync that's getting bigger over time, equally in both directions.

[Lio]

The music in question is this. The sweep isn't in the source files. The image up there is a spectrogram version of this:



I'd like to use phase inversion to drown out the orchestra as much as possible and keep the shawm (the weird instrument at the beginning). The top track is from the La-La Land release of the score, which has the orchestra only and did not include the shawm diegetic music. The bottom track is from a podcast co-hosted by a conductor for "Live to Projection" performances of the score, who used the audio stems from the film to represent the cue (which is what the YouTube video has).

If it is a speed issue, then what might have happened? Could it be it a sample rate thing? The lossless La-La Land file is only one or technically zero generations away from the source CD, but who knows how many times the podcast audio was encoded (probably from a 48khz source) so I guess that's the culprit.

[deleted user]

I see. Unless the two masters are identical, I wouldn't put too much hope into this approach. Even if you EQ them to sound the same, that filtering will cause phase differences I think.

With that said, you can probably calculate the drift by editing/calculating not based on time, but based on individual samples. For example soundforge can be set to operate in samples. You find two samples that match, in the beginning and in the end ideally, and calculate the amount of the speed difference.

Personally, if this was something I was obsessed with doing, I'd probably try spectral painting to isolate that instrument, as it seems to have a rather distinctive look on the spectrum. But the result likely would be far from perfect too, and it could potentially mean lots of work.