Free AI STEM splitting-MAKES ME APPRECIATE THE SOUND GUYS !!

Somewhere on the forum or an interview I remember Buck saying that in making Ghost Stories they used AI to "de-mix" George's stereo (I presume) tracks into some number of component pieces (like drums, vocals, etc.) for "re-mixing". Not being in the business I had a hard time believing that the de-mix results could be very good. So I had to hear for myself. I was surprised to say the least!

This had to be a zero-$ endeavor, so I prowled around the net. I found a number of on-line "Stem Splitters" where you could upload a song and get pieces (stems) back. One or two were free, others $ to $$$$. But I don't like creating accounts and uploading ANYTHING in general. I wanted a completely local solution. Then I found this:

Audacity is a free audio editing program for Windows that I've been using off and on for years. No expert at it since it keeps getting more complicated, but I've used it enough over the years to be fairly comfortable with the terms, interface, capabilities, and general use. If one is starting from scratch there is a pretty significant learning curve.

So I followed their guides and managed to get the "OpenVINO™ AI Plugins" installed. Not too bad to do. But the file sizes are huge - around 2 GB, bringing the Audacity directory up to about 3 GB with all the other add-ons I already had.

Now time to try it out, so I had an unnamed BOC pair of tunes in one file laying around that I used. Buck sings on both tunes. After I got the file imported, time to run the AI! I ended up with the following:

It only broke the original into 4 categories: drums, bass, other instruments, and vocals. Yea, 4 things, but remember it's FREE.

Finally, time to play with my new toy. Listened to the original stereo track to see what it sounded like. Then muted it and played all the pieces at once. I was surprised at how closely the mixed pieces sounded like the original, in fact I couldn't hear any difference. A trained person with headphones or studio setup probably could. But PLENTY good enough for me!

Then I started playing just one piece at a time. Really "strange" sounding (to me) when playing just one at a time. Selecting all the pieces but vocals I could hear NO vocals at all. Really cool.

Then the test I was waiting for - just the vocals. JUST BUCK plus the expected low-volume background voices. Way cool! Really weird to hear him sing the whole tunes with large gaps in-between that were, you guess, the drums+bass+other instruments. Not sure how well his voice got broken out, but a lot of variables that could have affected quality, the AI included. But still really good! Makes a great way to listen to a song you don't have lyrics to since all you get is the vocals. Like I said earlier, mix the vocals with the other pieces and essentially you get the original.

I tried playing with different combos of levels for the drums, bass, etc. Never got anything any better than the original.

Killed the day doing this, but I was stuck waiting in here all day for a phone call anyway.

Bottom line: gave me a toy to play with and MAKES ME REALLY APPRECIATE ALL THE WORK THAT GOES INTO RE-MIXING SOMETHING ALREADY MIXED. Yea, the pro tools are certainly better, BUT STILL!

FYI: I will NOT be releasing or posting any convoluted versions of the tunes. I feel that would be just plain disrespectful.