What's new
By:
Filters
LiteRECORDS

Register a free account today to become a member! Once signed in, you'll be able to participate on this site by adding your own topics and posts, as well as connect with other members through your own private inbox!

  • Guest, before your account can be reviewed you must click the activation link sent to your email account. Please ensure you check your junk folders.
    If you do not see the link after 24 hours please open a support ticket.

Powerful New Vocal Remover AI - Instructions

INGSOC said:
What is the best model to choose for instrumentals since there's 3?
Click to expand...

MultiGenreModelHP.pth - Better for pop songs and high pitch vocals.
MultiGenreModelNP.pth - Better with everything else.
StackedPart2.pth - Only run instrumentals created from any non-stacked model through this one. Might require more passes to remove vocal residue depending on the track.
 
***UPDATE***

Lite-records exclusive! This is a "sequel" sort of speak to the main Multi-Genre model released a few months ago. I haven't released this to my GitHub yet because I'm going to try for even better results and I need to come out with a new accompaniing stacked model.

This is by far the best model I've ever trained up until this point. Please let me know what you all think!

Multi-Genre Model 2 - click here
 
The new model is really outperforming every other one. Keep up the good work! I will update it on Colab soon :)
 
djtayz said:
The new model is really outperforming every other one. Keep up the good work! I will update it on Colab soon :)
Click to expand...

I agree. It's the biggest single leap in quality I've seen on any model to this point. At least on every track I've tested with it so far.
 
I want to thank you all here for the encouragement, help, and feedback! This whole project is a testament to what multiple people with a shared goal can accomplish!

It's crazy to think none of this would have been possible without the help of individuals all over the world, from the USA, Canada, Austria, United Kingdom, Russia, Japan, and Brazil (probably more) joined only by our shared love for music!

More models will follow very soon!
 
Thanks for all of this!
One quick question: Will the old stacked models work at all with the new model, or should I wait until you release the new stacked model before I jump down that rabbit hole?
 
Pants said:
Thanks for all of this!
One quick question: Will the old stacked models work at all with the new model, or should I wait until you release the new stacked model before I jump down that rabbit hole?
Click to expand...

You can still use it. It just might not be as effective because the newest model was trained with different parameters.
 
Wow the new model is outstanding! :*
If this can get even better some tracks will sound pretty much official :D
Thank you so much!
 
chilinvilin said:
Any way you could get the current version on colab or is there something I could do to get the new version on colab from my end? Thanks
Click to expand...

I just updated Colab and you should be able to run the new multi-genre model, alongside the older ones too.
 
I know for most having the cleanest instrumentals possible is a must. But for some of us it is important to have backing vocals.
Be it may I don't know much about coding, I do have a little understanding of how some of this works. The models for the most part use clean vocal tracks and clean instrumentals to help train the models. Unfortunately the songs I have uploaded for [MENTION=39673]Anjok[/MENTION] that are Evanescence from Rockband do have backing vocals in the vocal tracks. I feel using those to try and train a separate model will help the A.I. pull more backing vocals out of the instrumental in the pella track which will allow me to filter out the backing vocals better for my use. As it stands the A.I. gets most of the pella but also bits and pieces of the backing vocals which is why I can't get good filters. I do hope something is possible because there are still many Evanescence songs that never made it out to karaoke venues and I want to make the best renditions of them possible. Thanks. I'm referring to the Google Colab method of course since it's possible to add more than one model to use.
 
Last edited:
i am but i don't see that. also, could be a cool idea to have a checkbox to make save to the same location as music file location? just to save some navigation time :) happy i got the GUI working. it's dope!
 
NewAgeRipper said:
I know for most having the cleanest instrumentals possible is a must. But for some of us it is important to have backing vocals.
Click to expand...

I feel the backing vocal pain. I do wonder what the best approach is. I have found most every model to date has tended to leave choral type backings in at least mostly, so I think that making a custom model for that type of backing might be very much a future option. Granted I know nothing about the actual training process. Anjok is the expert here on that. Just an observation.

One thing that has been extremely useful for me, even though it is time consuming, is spectral editing. If you haven't dug into it, I highly recommend it. Not only is it critical for cleaning up any left over bits and pieces, but it can also be a big help with backings. Specifically if the backings kind just overlap the main vocals a little bit, often times I can isolate and clean out one of the two layers effectively. Like if one line is just holding a note, and the next one comes in with a hard consonant sound like S or T, that sound represents as a thick vertical line across the spectrum that can be muted, while the long held note is horizontal. Easy. Otherwise, so long as the note is different on the next line, you can usually identify it because it will start in a part of the spectrum that the previous note was not occupying, so by isolating and muting out those horizontal lines, you can pull the vocal layers apart as well.

Not the easiest solution of course, but I have managed to use it effectively in many cases so far. Course when it's extra complicated, like two different vocal lines on top of each other completely, yeah, I do get the feeling that trying to train the AI to do one and not the other might be next levels of challenging.
 
The problem with spectral editing is taking out something usually takes something important in the same spectrum with it. It's always best to do minimal surgery for best high quality results in my opinion.
Still waiting for the vocal making side to catch up. Long ways to go. Pretty much all of the UVR vocals are washed out and unusable.
 
INGSOC said:
The problem with spectral editing is taking out something usually takes something important in the same spectrum with it. It's always best to do minimal surgery for best high quality results in my opinion.
Still waiting for the vocal making side to catch up. Long ways to go. Pretty much all of the UVR vocals are washed out and unusable.
Click to expand...

In isolation yes, but when the vocals are mixed in with any music, they sound more solid like you expect them to be. I've muted and unmuted the vocals while playing both them and the instrumental listening for that difference. I'm not saying it improves the vocal track. Just saying it hides the washed out sound and gives the illusion of sounding better as long as there is music present.
 
You must log in or register to reply here.