A few years back I took an interest in the recent phenomenon of AI stem isolators. Installing the likes of Spleeter and later Demucs. Initially I wasn't too impressed as I expected, with the usual AI artifacts, like muffled filtering, waving in and out and noise as it struggles with full on sections.
I found that I had few songs I had isolated and forgot about. I did start to use these stems for some purpose. And that was as a reference for clean stems I had for volume balancing.
But lately I tried some recent ones. I thought I would have a go at UVR, which actually has been around for a while, and tested it online. I was impressed by the results. I gave it "Billy don't you lose my number" and I could barely fault the instrumental or vocal.
Then I was reading in the chat window some comment about x-minus.pro and MVSep. Turned out I already had browser tabs open for each. And with UVR linked to x-minus. Well it inspired me to look again and then I became obsessed for a few days testing out all these songs. Oh wait I haven't stopped lol.
Things look they are improving massively. I've tried a few songs and some came out clean as. Others still need work. What I found is that synth pop and synth rock tends to be cleaner but "real" recordings like rock are lower quality. It's able to isolate some guitars and piano quite clearly. But it can struggle with drums. It seems a bit hit and miss at times, as I find the kicks sound sucked out, the snare is better but sometimes muffed, and the cymbals which I thought would be more complex sounded clear. :-D
So right now there's about 3 levels we have:
I don't know what other peoples work flow are but I would prefer a nicely balanced stem I can work with. I'm not a studio professional for the last 20 years. So needing to mix and master is a skill above my creative ideas. My use case is for extended dance remixes since it's what I grew up with. Sometimes I have an idea for some DRT, that is some drum replacement therapy, but mostly my ideas which go back decades are rearranging the original and laying FX on top. Doing the creative work is time consuming enough for me. so although final mixing and mastering is important to the process, I would prefer to spend time on the arrangement so I can share my ideas out there.
The AI approach may assist here. Even though there's a likely quality loss it tells you what it should sound like. Perhaps some hybrid approach could help by using an isolated source to assist AI in extracting a stem.
What do you think, will AI stem isolators become the game changer?
I found that I had few songs I had isolated and forgot about. I did start to use these stems for some purpose. And that was as a reference for clean stems I had for volume balancing.
But lately I tried some recent ones. I thought I would have a go at UVR, which actually has been around for a while, and tested it online. I was impressed by the results. I gave it "Billy don't you lose my number" and I could barely fault the instrumental or vocal.
Then I was reading in the chat window some comment about x-minus.pro and MVSep. Turned out I already had browser tabs open for each. And with UVR linked to x-minus. Well it inspired me to look again and then I became obsessed for a few days testing out all these songs. Oh wait I haven't stopped lol.
Things look they are improving massively. I've tried a few songs and some came out clean as. Others still need work. What I found is that synth pop and synth rock tends to be cleaner but "real" recordings like rock are lower quality. It's able to isolate some guitars and piano quite clearly. But it can struggle with drums. It seems a bit hit and miss at times, as I find the kicks sound sucked out, the snare is better but sometimes muffed, and the cymbals which I thought would be more complex sounded clear. :-D
So right now there's about 3 levels we have:
- Stems typically sourced from games.
- Multi-tracks from official or actual song.
- Isolators using AI.
- Stems tend to be mixed but lack balancing and final mastering.
- Multi-tracks tends to be raw and include extra takes.
- Isolations give you a mastered stem but all mixed can still be too loud.
- Stems tend have the lowest quality for the convenience.
- Multi-tracks are usually the best in studio quality with some in lossless 24 bit.
- Isolations could in theory match source but suffer from quality loss in filtering process.
I don't know what other peoples work flow are but I would prefer a nicely balanced stem I can work with. I'm not a studio professional for the last 20 years. So needing to mix and master is a skill above my creative ideas. My use case is for extended dance remixes since it's what I grew up with. Sometimes I have an idea for some DRT, that is some drum replacement therapy, but mostly my ideas which go back decades are rearranging the original and laying FX on top. Doing the creative work is time consuming enough for me. so although final mixing and mastering is important to the process, I would prefer to spend time on the arrangement so I can share my ideas out there.
The AI approach may assist here. Even though there's a likely quality loss it tells you what it should sound like. Perhaps some hybrid approach could help by using an isolated source to assist AI in extracting a stem.
What do you think, will AI stem isolators become the game changer?
Last edited: