Beyond Rips With Audio Seperation

[moodmusic]

I've been involved with emerging AI audio technology. For those not up to speed, here's what it can do.

James Brown - Escap-ism - Vocals
https://mega.nz/file/6NpwSSLa#c_s5OgpyNlMWMVhk32rFA4UabRF10qyDwAJxskvchQY

The difference with this AI model from what's publicly out there, is the quality level. Yeah, record labels are very worried about it.

[Retronic]

I’m a fan of these Stem rippers and use Izoptope ‘Music Rebalance’ and have stripped a few library tracks back.  Great sampling potential and just fun to mess around with.

[tonyc1971]

I believe this tech is being used alongside Apple Music for its ‘Sing’ section of music in their libraries. Essentially within the music app, it offers the user the choice of reducing the volume of the lyrics so you can sing along to the song. Similar to how Izotope works I believe.  Clever stuff !

[Retronic]

I did a comp of some random tracks I’d stripped the vocal out of a while back: retroteque.wordpress.com/2021/05/01/no-words-1-instrumentals/

[moodmusic]

Izoptope is ok, but very fragmented. The AI model I'm working with is far more impressive at conversions because it's trained on more music. I run the code so I can do batch conversions of entire albums easily.

This is still the 1st generation of AI audio. Give it two years and it will be master quality. It will even restore audio, similar to how old black and white images/videos are restored to HD.

Some more...

David Axelrod - The Edge - Instruments
https://mega.nz/file/qQJDVKpT#mainkm3Ktof1o6ddiam2-hZ0ou4MszQOtuZG5ouZmy8

Gil Scott Heron - We Almost Lost Detroit - Drums
https://mega.nz/file/KIpnRKib#6NCOtExZjRX2ICe0T0ZiRMwk8A23gYXR21O47dmGxp0

[potzorbie]

I get an error trying to open the links

[likedeeler]

I've been involved with emerging AI audio technology.
... record labels are very worried about it.

Why are they worried? What exactly do you mean by "AI" in this context?

[nidostar]

Because of my oriental connection I've been into karaoke for more than 40 years now. But I'd always been disappointed with the quality and standard of the average karaoke backing track. I've spent many hours with SoundForge, Audacity and other audio apps trying to isolate the vocal from the rest of the track but without success. I remember back in the early 90s having a Sony CD player that used a simple algorithm for cancelling out the vocals at the push of a button but only if the vocal was dead centre. But this is something else entirely since I take it you can isolate more than just the vocals? I'd love to know more though sadly, as with potzorbie, I get an error when trying to access the links.

[Greta]

Don't have a clue what AI is but is certainly not "Vintage Library vinyl and CD rips".

Moved to "General Discussion".

[nidostar]

I did a comp of some random tracks I’d stripped the vocal out of a while back: retroteque.wordpress.com/2021/05/01/no-words-1-instrumentals/

Ah so that's how you produced the instrumentals. I thought they were genuine backing tracks that you had somehow managed to get hold of. I've really enjoyed listening and singing along to those tracks. But Izotope doesn't come cheap. Impressive.

[moodmusic]

@potzorbie Not sure. What's the best free site to share mp3s?

@likedeeler AI as in artificial intelligence for music. The labels are worried b/c this technology will devalue their masters. It's probably why they started offering their music catalogs for sampling on sites like Tracklib and Usample. They best make money while they can. Aside from stem separation, AI will be able to create songs on the fly. It's crazy.


@nidostar This is much different technology, not phase cancellation with vocals. If you want to try a song, post a link to the file and I'll separate it for you.

@Greta Sorry, I'm new here.

[Lord Thames]

I've used AI to create stereo-ish mixes of a couple of KPM tracks that should be in stereo but aren't - I suspect the technology has improved since I did them, though, and it might be fun to have another go.

Is the system you're using a secret, moodmusic?  I love trying stuff like this! I might put up a couple of tracks for you, if you don't mind

[Fuzi]

Welcome @moodmusic! This is a pretty exciting field indeed. I am always excited to read about AI, tho I care more about AI as a tool than AI as a creator, since I've followed progress on the latter and honestly didn't find anything worth looking at.

Wasn't able to access the samples you shared on file.

mega.nz is a good alternative I like to use.

[stackjackson]

I get an error trying to open the links

Yeah, none of those files are available. Best never to 'hotlink' the urls.

I'd love to hear these samples.

[moodmusic]

@stackjackson @Fuzi I fixed the links to MEGA.

@Lord Thames - It could be done with AI if the model was trained to do so. Mono to stereo is tricky. If you want a natural sounding stereo mix, the best way would be to mic up a pair of speakers in stereo. That's what Bruce Swedien did for Micheal Jackson's albums. He would take a mono signal for something like a Moog synth and play it off the studio room speakers — then record the speakers in stereo with a pair condensor mics. I've done it before. It works. But you need really good studio equipment.

It's no big secret. It's just pieces of code that data scientists are using for AI research. There's some paid services like AudioShake and Spleeter, and free code such as Open-Unmix, but they don't sound as good.

Although I should mention that the quality of the recording matters. From my batch runs I've noticed that anything recorded before the mid 70s doesn't convert well. Likely because they were recorded on 4-8 tracks with a low ips tape speed (more saturation noise). For 60s music 1-3 tracks. Anything from the late 70s onwards starts sounding really good, depending on how much instrumentation there is.

Post up some tracks. I'll run it on my next batch.