Various versions of the same song
Sometimes the same track you may try to isolate can exist in few versions: e.g.
0) album version
- on streaming services - sometimes both explicit and non-explicit album versions are available, plus sometimes both in 44kHz 16-24 bit or 48-192kHz - they might give a bit of different results (if your old player app struggles with playing these files, download the latest MKVToolnix, drag and drop the file and begin multiplexing. It doesn’t reencode/recompress the audio stream)
- on CD - sometimes these are two different masters - if total time and track gain of lossless files scanned by F2K is different by e.g. 1dB or iLufs reading is different, it’s a different master (both mp3 and lossless file should show exactly the same iLUFS (to 0.0X precision) if they’re from the same source, if they’re not - the changes should be noticeably on spectrograms comparing both lossless sources - allflac has only CD rips for at least pre-streaming/~pre-2016 era albums).
~ Sometimes recent masters of older music are louder on CDs than on streaming services providing fewer dynamics, and in most cases such CD should be worse for AI separation when mastered to -9 iLUFS vs -14 iLUFS for streaming, although it can be totally opposite for some releases too
~ Regional CD version - certain albums in the past used to have different releases for some countries, e.g. Japan, different track order, even slightly different mastering
- on DSD - if available, they are different masters and might be worth to check for separation too
- on SACD - -||-
- on vinyl (so-called “vinyl rip”) - might give you a bit of different results for problematic tracks
- on DVD-Audio (sometimes also 2.0 releases) if AC3 was used, it can be lossy - used bit depth or sample rate might depend on a release and it can be a different master
in separation - usually vinyls are different masters with bass more/mostly in mono
1) single version - in the old days, single versions contained official instrumentals or accapella which not always invert with original mixture to get instrumental if it wasn’t available (but if it inverts at least partially it might give you better result - always try lossless files - lossy might not invert well)
- on CD - sometimes contain different track list than on vinyl, extra tracks, remixes, etc.
(rarely available on streamings in this form now) - always refer to Discogz to find all releases of your interest
- on vinyl - -||- (won’t invert correctly due to constant playback speed fluctuations)
2) deluxe edition/reissue/remastered (sometimes separated instrumentals from remastered versions can be crispier than leaked multitracks which are rarely even mastered; also, different remasters might be available on streaming platforms or fan-made ones on YT or on the internet)
3) Video released for the song - although lossy, it can be a completely different mix or master giving different results for separation (YT, Tidal, Vimeo, Dailymotion, etc.)
4) Release from movie OST or from a game (OST or so-called gamerip) - usually different master
5) Official remix - sometimes it might be easier to separate vocals from such version
6) Leaks of earlier version of the song (might have different mixing, lyrics, even instrumental) - YT (often a subject to be taken down), fan-made Discord servers, internet (use various search engines; more below)
7) Leaks of multitracks or stems of the song - usually it’s a different master than the final song, but you might experiment with Matchering and using it mixdown without vocals a target, and well sounding separation as input for Matchering to make it more similar to the final song
8) Leak of instrumental/vocals - it can be lossy or slightly different from the final song, e.g. close to final stage or sometimes might have even dry vocals without any effects
Surround versions
8) 5.1 - e.g. DTS on DVD Audio/Blu-ray or SACD (you can search Discogz to look for multichannel versions released on disks, e.g. entire DTS Entertainment label)
9) 360 Reality or Atmos (7.1 or more) - e.g. on Tidal, Apple Music (better quality/more channels than Tidal; how to download described below).
Sometimes in surround releases, vocals can be there simply in the center channel, but it's not always the case - still, it can be a better source, e.g. when you manipulate with volume of specific channels, or for vocals - when you get only center channel with very little instrumentalization which may turn out easier to separate by AI (for instrumental you might possibly invert result of vocal model and center channel to receive the remain instrumentalization in center).
“For me, I convert left and right together then center alone then LS+Rs+LFE together then I have 3 audio files process them then remix into 5.1 again.
The 2 are in stereo and one mono which is center” - killdubo
“I do the same, except that I don't process LFE. Only the other 5.” - santilli_/Michael
E.g. "With the Dolby Atmos release of Random Access Memories, some vocals and instrumentals can be separated almost like stems"
Or alternatively, you can simply downmix everything to stereo and then separate (just to check the outcome vs regular 2.0 versions).
There are different approaches to downmix multichannel audio to stereo. E.g. you can use Foobar2000 and RBM on a file>Convert>Downmix to stereo option added in DSP Manager and convert to a lossless format (potentially in 24-bit or higher). Watch out for clipping - you can apply a limiter with foo_dsp_vst3 (which also supports VST2) or export it as a 32-bit float so the volume can be turned down without clipped peaks in your DAW after the export from F2K, and there use the limiter.
Also this librosa script (for 5.1 or 7.1 inputs) might give a bit different results during downmixing.
And also various ffmpeg scripts exist (e.g. a thread with collection of them), varying at least from the F2K method sonically (e.g. Dave_750's method script - edit paths; volume parameter accepts negative values; or read more, and here).
You can also use tebasuna51’s method there for the FreeSurround Decoder plugin for F2K set like here.
The best sounding method might depend on your specific songs or your needs.
If you want to use commands to increase volume during the downmix, control the track gain value afterwards when you scan replay gain values in Foobar2000. Otherwise you have clipping. Keep track gain not exceeding 1.00, otherwise you have clipping even with still relatively quiet music, but louder than any channel in the original mixture (not always a big issue - it commonly exists e.g. after reencoding to a lossy format.
Tape
10) Cassette tape rip - wouldn’t recommend it as a source (maybe unless it sounds superb, or you have a great deck at disposal to rip the recording) - even though the cassette tapes might be still released occasionally, contemporary music usually sound worse on them than it used to in the past, and potentially to compensate for it, they might contain different masters (cassette tape won’t invert due to speed fluctuations either)
11) Reel to reel tape rip (better quality, mostly old music, also, usually in the past, the base medium for archival original stems of the recording before final mix and master, but degrading along the time, sometimes can even change the sound after some time during even the period of the song production/finalization/final mixing/before transfer)
General
As for a good quality music on streaming services, you can get stereo FLAC 16 bit and 24 bit on Qobuz, Amazon or on Tidal (now also up to 24 bit FLACs for Master quality - formerly MQA 24 bit (in the past most of Max (formerly Master) quality on Tidal was 16 bit MQA, while High (formerly Hi-Fi) is and was always FLAC 16 bit; MQA was lossy (but less than all other formats), but 24 bit MQA file could have given better results than 16 bit FLAC). MQA was gradually changed to FLAC on Tidal, but seems like old uploads in 24 bit MQA are not 24 bit anymore, but just 16 bit FLACs, so you might want to use Qobuz to find some of these 24 bit files if they aren’t available on Tidal like they used to be.
Most importantly -
Feel free to experiment with different versions and find the best result with a specific version of your song, although 24 bit FLAC should be the best (although not everyone might notice the difference).
If you have seemingly the same FLAC Audio CD rip from before streaming services times (~<2013), it can happen that a lossless file taken from a streaming service may be slightly different in most cases (same length but slight changes in Spek across the whole track which normally don’t exist when comparing FLACs from various streaming services which have the same Audio MD5 checksums - also sometimes track finishes in slightly different place). Sometimes it can sound better, sometimes worse.
([outdated - there are no longer MQA files on Tidal] and we’re talking about situation that it’s not MQA 16 bit like "Master" quality files on Tidal [but lots are 24 bit as well, though it's better to get them from e.g. Qobuz if 24 bit for some track is available, or at least compare both, because it can give slightly different results/quality]).
Also, it can happen that 24 bit MQA on Tidal will sound better for whatever reason than seemingly better FLAC on Qobuz - it might be possibly due to different files sent to streaming services by the provider/label.
How to notice difference on spectrogram in e.g. Spek between MQA and FLAC is frequencies from 18kHz (only in certain places) but in all cases - frequencies from 21kHz - press alt-tab between the two windows’ - don’t hover your mouse between preview of both windows’ - use alt-tab - you’ll notice the changes easier. That way, you’ll notice CD rip vs streaming differences if there are any.
Generally, MQA is the least of lossy codecs - you might consider picking it where its 24 bit variant is available over regular 16 bit FLAC (separate the track using the two, and you should notice any differences easier if you already can’t hear them on mixtures/original songs).
Comparisons of various versions of the FLAC files on streaming services
Use Audio MD5 in Foobar >properties of the file (or download AudioMD5Checker) to not run in circles looking for various versions of the same track with the same length. Some FLAC files don’t have MD5 checksums in F2K shown, so you’ll need to download AudioMD5Checker.
E.g. on Tidal Recovery by Eminem returns the same MD5 for Deluxe and regular album, but using Deezer source, checksums are different for both (to differentiate - albums on the net sites have various release dates), but Deluxe on Tidal with the regular on Deezer have the same MD5. And when Audio MD5 checksums were different, there were different results after separation. In this case of one unique vs 3 same MD5, the unique resulted in worse separation (but it can depend on more factors in other cases). Be aware of non-explicit versions which will naturally have different checksum. Some FLACs will lack MD5 calculated, so you will need to use AudioMD5Checker above, or just look at spectrograms in e.g. Spek.
Usually if both compared files differ in iLUFs or Track Gain reading from F2K, and esp. the song length is different, it’s a different master, or at least reencode.