Most of My Catalog Only Exists as a Stereo Bounce
I have more than fourteen thousand tracks in my catalog, and the honest truth is that a big chunk of the early ones only survive as a stereo mixdown. Two channels. Left and right, already summed, already committed. No stems, no session file, no way to reach back in and grab the vocal by itself. Some of those sessions lived on a drive that failed. Some were bounced on software I no longer own a license for. Some I made before I understood that you should archive stems like they are money, because they are.
For a long time I treated those tracks as frozen. Whatever they sounded like on the day I bounced them was what they would sound like forever. If the vocal was harsh, it stayed harsh. If the whole thing was too quiet next to a 2026 release, it stayed quiet and got skipped. That is a bad place to leave a catalog you are still trying to grow, because a new listener does not care that a song is eight years old. They care that it sounds thin next to the track that played before it.
Why You Can't Just Put a Limiter On an Old Bounce
The instinct is to slap a limiter on the stereo file, push it up, and call it a remaster. I did that for years and it never actually fixed anything. When everything is already summed into two channels, every problem is baked in together. If the vocal is too bright and the low end is muddy, there is no move that helps one without hurting the other. Turn up the highs to open the vocal and the cymbals turn to glass. Roll off the top to tame the harshness and the whole record goes dull. You are negotiating with a mix that already made its decisions.
Loudness makes it worse. Pushing a bad balance louder does not make it sound better, it makes it sound louder and worse. And if you overshoot, streaming platforms turn you back down anyway, so you end up with a squashed master that gets normalized to the same level as a clean one — except yours lost its transients to get there. You cannot master your way out of a mix problem. You have to get back inside the mix.
The Recovery Move: Separate the Bounce Back Into Stems
This is where source separation changed what is possible. I run BS-RoFormer on my Mac, on the MPS backend, to take a finished stereo mixdown and pull it back apart into stems — at minimum a vocal and an instrumental, and depending on the track, further than that. It is not magic and it is not lossless, but it is good enough that I can suddenly touch the vocal without touching the beat, and touch the beat without touching the vocal. That single capability is the whole game.
Once I have a recovered vocal sitting on its own track and a recovered instrumental on another, the old frozen bounce becomes a session again. Now the harsh vocal is a problem I can actually solve, because it is isolated. Now the muddy low end lives on the instrumental where I can address it without smearing the words. I got the leverage back that I lost when the stems disappeared.
Repair Only What's Broken — Never Re-Produce
The discipline here matters more than the tools. The temptation, once you can touch everything, is to remix the record — re-EQ the whole thing, add new elements, "improve" it. That is a trap. I am not trying to make a different song. I am trying to make the same song compete. So the rule is repair-only.
On the recovered vocal I do the smallest amount that fixes the actual defect: de-ess the sibilance that always got worse in separation, gently de-noise the artifacts the separation introduced, tame the one or two frequencies that were fatiguing. I leave the performance alone. On the instrumental I usually do even less — often nothing, because the beat was fine, it was the sum that was fighting itself. The goal is to reinject only the correction, not to repaint the picture. If I catch myself reaching for a creative move instead of a corrective one, I stop. The version of me that made the song gets the final say on how it should feel; the version of me remastering it is only allowed to clean the glass.
Re-Master to −14 LUFS, Matched to a Reference
Once the vocal and instrumental are repaired, I sum them back down and run the result through my mastering pipeline — reference_master.py, the same never-regress workflow I use on new records. It matches the remaster to a reference track I trust, and it lands at −14 LUFS integrated, which is the level that survives streaming normalization without getting crushed. The "never-regress" part means the pipeline will not let a new pass come out worse than the last one; loudness only moves toward the target, not past it into the squashed zone.
The result is a version of an eight-year-old song that can sit in a playlist next to something I finished last week and not embarrass itself. Same performance, same intent, just no longer thin and no longer harsh. For a catalog this deep, that is enormous — every recovered track is a track that can go back into rotation instead of getting skipped.
The Honest Limits
Separation is not free of cost. Push it too hard and you hear the artifacts: a slightly watery vocal, a smeared transient, a cymbal that shimmers wrong. On some tracks the separation is clean enough that nobody would ever know. On others it is worse than the original problem, and the right call is to leave the song alone. Not every record needs saving, and a remaster that trades one flaw for a stranger one is not a win.
I do not have a GPU render farm for this — it runs on the Mac, on MPS, one track at a time when I decide a track is worth it. That constraint is actually good, because it forces me to be selective instead of running the whole catalog through a blender. I remaster the songs that are getting played and deserve to sound current, and I let the deep cuts stay as they are. Good enough to compete on a playlist beats perfect in a folder nobody opens.
FAQ
Can you really recover usable stems from a finished, mixed-down song?
Yes, within limits. AI source separation like BS-RoFormer splits a stereo bounce back into a vocal and an instrumental (and sometimes further). It is not the original multitrack and it is not lossless, but it is clean enough to let you repair one element without touching the others — which is the entire reason to do it.
Doesn't AI separation degrade the audio quality?
It introduces artifacts, so you have to be disciplined. The trick is repair-only: fix the specific defect on the isolated stem, de-ess and gently de-noise, and reinject just the correction rather than re-producing the track. On a clean separation nobody notices; on a bad one, the right move is to leave the song alone.
What loudness should I target when remastering an old record?
−14 LUFS integrated. That is the level streaming platforms normalize toward, so mastering louder just gets you turned back down with worse transients. Match the remaster to a modern reference and land at −14 so the old track sits next to current releases instead of under them.
Do I need an expensive GPU to do this?
No. I run the whole thing on a Mac using the MPS backend, one track at a time. It is slower than a dedicated GPU box, but that constraint keeps you selective — you remaster the songs that are actually getting played, not the entire catalog on autopilot.