Where is the download for this? ;-0

 

can someone explain to me what stems are?

 

The "slices" of the music, Generally this type of software are able to slice the music by instruments/instruments bus !

 

Full track length renders of groups, not individual unmixed multi-tracks.

 

Huh? Under you name it says "Producer" and you don't know what stems are?

 

You might be surprised how often people get it wrong. And also how often people will ask for a "stems mix" quoted price, agree to it; and then try to send a mulitrack session, as if they don't already know the different amount of work involved for each.

 

I doubt that you are either highly trained in countrapuntal art...

 

Glenn Miller as both a producer and kapellmeister as early as 1941 was brilliantly able to separate stems by using only farm machinery of the times. His success was great, and using primitive midi technology was able to trigger his bands instruments to play their parts. The Nazi's coveted his work, captured him when they shot down his plane flying over the English Channel on Dec. 16, 1944, tortured him until he gave away his secrets and used that knowledge to imitate his music with Goebbels' propaganda swing band, Charlie And His Orchestra.

And so now you know.

 

individual tracks of instrument busses, so for ex. all guitars within one track, whole drum kit within a track, all vocals within a track

 

for drums isolations, just use UVR5 with ht demucs v4 model (downloads though UVR5 settings)

 

Since there is really no comprehensive ensemble model right now that yields the best results of all current bleeding-edge technology on each stem, I've started doing reductive passes on the same file, whittling it down with different models one instrument at a time. Target Instrument / Other. Then run the "Other" file through the next stem model in succession.
It's kind of a bummer that there is no Mel or BS Roformer full ensemble model that hands-down beats all other models.

IMO right now the best models are UVR-online's MelRoformer Drums, MVSEP.com's BSRofomer Vocals and BSRoformer Bass.
There is not really a good Guitar model in existence yet, which is kind of disappointing. So lately I have just been extracting the other 3 stems reductively and the final remainder is the Guitar stem.

Also, depending on the quality of the source material, sometimes older models like last years MVSEP-MDX2023 ensemble can give better results than MelRofomer on Drums if the input material is for example cassette-tape sourced or post-FM, I believe this is because MelRoformer hasnt been as extensively trained as HT-Demucs for example. Models only trained on A+ quality material have issues with separating lower-quality source material. I'm probably in the extreme minority for that usage scenario, as I do restorations of less-than-perfect source material.

Furthermore I am a bit disappointed that there isnt an Audiophile quality-checker giving input on some of these new models.
What I mean is, while new models may score impressive SDR metrics, when you check the spectrograms of outputted stems, some of them are riddled with Junk Frequencies, usually horizontal bands of bullshit frequencies that are not in the source material. This seems to be a common issue with MDX architectures. So I wish these engineers would confer with, or maybe learn themselves, that SDR metrics arent the only qualifier of whether a stem model is actually good. Frankly at this point no developers should be releasing new models that have this issue. The new SCNet-Large models are afflicted with this issue.

To see what I am talking about just load the stems up in an editor that has spectral display, or use the program Spek to do a quick view of any stems.

While maybe this wont affect most people, stems with garbage frequencies imparted by AI models, will affect the overall quality whether it's audibly discernable or not.

 

Unfortunately no, the terminoogy is very confusing. I'm only a "producer" here because I post a good bit.

 

Dang that's fucking awesome, it seems like a way easier task to remove vocals than 20 years ago..

 

Has anyone tried Moises?
Looks like they have acoustic/electric and lead/rythm guitar separation models, and their own "DrumSep"
Not sure there are any guitar further refinement separation models out there on the web but their demo looks pretty astonishing if it actually works that well.

Saw someone on reddit claiming Moises bass/guitar/drum models may be better than UVR or MVSEP (I'll believe it when I hear it)
Gotta have a "Pro" account to access them ($20/month)..

 

All uvr models I've tried (all that jarredou had on his collab) struggle with the slap bass.. Slap sounds (which contains mid to high fq) goes into guitar stem, so that's still hard

I check it on this track

 

Nice track.

What rythm pattern is that ?

 

If anybody has a Moises Pro account, please PM me, I'd like to send you a couple test files to see if the rythm/lead and acoustic/electric separation is worth paying $20/month for

 

one beat 7/8, other 6/8. so uhh 13/16?

also, bout this track. i found it on a torrent tracker in "fusion-prog-altrock" genre, back then I didn't use streamings too much, and furthermore its was just a good way to get me some starting point. another band that struck me hard was "Pushing Daisies Band", with their debute album

. shit dude, I play these melodies in my head and I get all covered in goosebumps... it was also that incredible feeling when you find something really special, that nobody seems to know.

but really soon manually browsing became too tiresome. and i had this "Bad Suit" record, which is pretty unique and I was unable to find anything similar or at least isomorphic enough to put it into a virtual dialogue, you know, to create a virtual discourse, a sub-genre in my own head for musicians to polemicize

that's when I decided to try Chosic AI, which suggests you music based on a track you feed to it. you got some neat settings to tweak, like popularity setting, instrumental/vocal. and yeah, I found some bands to put them together with this discovery. they never hit me as hard though, but I discovered a good range of bands, which still listen from time to time. like Marbin, Shawn Lane, Mark Lettieri, Solar Wind, Tauk, AKINETON RETARD. There aren't too many of them. Maybe I should do it again some time, and also check new releases on trackers. These torrents man, even though we kinda steal from musicians (or as Gabe Newell put it - their product isn't good enough for us to buy it), are still a great way to find something new. Maybe there's lot of ways of discovering new music, AI being one of them, then we got personal advice (like when you read some blogger or critic who kinda introduces you to his personal World_Of_Sounds™ (aka personal user interface), and then you got this word-of-mouth via reddit, torrent trackers, etc.

 

Last edited: Jan 3, 2025

Which program can currently separate a track into the most instruments?