Peter Jackson and MAL audio question

[peter61]

I wasn't sure where to post it, if it's not in the right place, please tell me.


I'm working for a while with audio software to create the best DES possible and this implies separating instruments and vocals the best to my knowledge, of course, the results are not perfect, but I always work with the intention to get the best results possible. I mostly work with mono mixes and also with stereo mixes to separate some elements, with software like Demix pro, RipX and others.

First, here is the link to a snippet of a vocal stereo synch mix of "Day Tripper" which comes from the Peter Jackson You Tube video that someone extracted lately:

hxxps://mega.nz/file/kk0kBKbZ#3e8QtC8lE ... T_H268nTzw

So, here is my question:

About MAL ( and also perhaps Dolby Atmos)

Until now, I have never heard such a perfection from an extracted element like this snippet (and don't forget that the audio source is from a lossy You Tube video).

Considering that this song has been recorded from a 4 track machine in 1965 and that the vocals should have been recorded from the same track (mono), how they were able to get such a good result like this?

I'm simply curious to know how they did that, if it's not a "military secret" (because, it could be..)

[Tex]

They sent PJ a copy of the original 4-track to start with.

The vocals are on two of those tracks. One vocal track included the tamborine and lead guitar solo or the solo was maybe punched in the other vocal track not sure other people will know.

[BDJ]

The guitar was recorded on the same track as the tambourine and some vocals. This left another track with nothing but vocals. That last track would be an excellent 'feed' for MAL.

[Lord Reith]

I was most impressed by his demo and the extremely clean extractions. Going by what he said, I think this is almost entirely due to better and more training data. He says his guys had a drum kit set up to create their own training samples, and I would imagine they've recorded conributions from other musicians too to try and generate the perfect training set.

The actual technology seems very similar to Demucs, which is a "hybrid" algorithm that uses a combination of waveform and spectrogram-based approaches. I'm even hearing the same artefacts on the vocal extractions.

The closest thing to what he is using is MDX23, which gives very clean vocal separations (the best I've heard). For drums, Demucs4 gives probably the best results, although Demucs3B may give better results for some drums. For guitars and pianos and other instruments of that sort, Demucs3B gives the most consistent good result in my experience. And for bass guitar, it's a crap shoot if it's not strong and loud. For good loud bass lines, Demucs3B does a good job. For some of the bass lines it misses, Demucs4 may pick them up. And for bass lines which both of them fail to capture, UnmixXL is the fallback for me.

All these algorithms are available to use on mvsepDOTCOM

[20YearsAgo]

I was most impressed by his demo and the extremely clean extractions. Going by what he said, I think this is almost entirely due to better and more training data. He says his guys had a drum kit set up to create their own training samples, and I would imagine they've recorded conributions from other musicians too to try and generate the perfect training set.

By this, do you mean they've actually brought in musicians to mimic the Beatles' performances instrument-by-instrument? And through "learning" those parts, MAL was then able to better identify and extract the actual instrument-by-instrument parts from the Beatles' master tapes?

[Lord Reith]

I was most impressed by his demo and the extremely clean extractions. Going by what he said, I think this is almost entirely due to better and more training data. He says his guys had a drum kit set up to create their own training samples, and I would imagine they've recorded conributions from other musicians too to try and generate the perfect training set.

By this, do you mean they've actually brought in musicians to mimic the Beatles' performances instrument-by-instrument? And through "learning" those parts, MAL was then able to better identify and extract the actual instrument-by-instrument parts from the Beatles' master tapes?

No, not really. For demixing algorithms they need to create a "training set' which is a bunch of short audio clips that represent what instruments sound like on their own and in an ensemble. It is very difficult to obtain clean quality sounds. The Beatles' multitracks are probably not much use because they contain instruments blended onto one channel. Or even with drums, there aren't many instances where you can hear Ringo just play one of his drums by itself. The only examples I can think of is the end of Please Mr Postman where I once managed to pinch kick and snare samples for a demo of One And One Is Two, and the German All My Loving where his lone hi-hat opens the song. So having musicians create a library of examples would be extremly useful to them. Having the sound of an isolated snare next to the sound of a snare as part of a kit teaches the software what to look for.

The instrument samples used in things like demucs are obtained from stuff on the web which is probably not ideal and not fine tuned to what might be required in some instances.

[lennonology]

Sorry for the long link, but I wouldn't be surprised if Waves' Abbey Road plugins were implemented as well.

CM

https://www.waves.com/bundles/abbey-roa ... gKyKPD_BwE