UVR5 the Best AI stem separation algo?

Discussion in 'Software' started by curtified, Feb 27, 2023.

  1. poly

    poly Platinum Record

    Joined:
    Sep 29, 2016
    Messages:
    285
    Likes Received:
    173
    Location:
    Hä?
    I didn't understand where i can set this chain in UVR. Models already downloaded.

    [​IMG]
     
  2. jarredou

    jarredou Guest

    First, you have to download all these models from UVR => Options => Download Center (you can DL only one at a time)
    Then, you have to choose the "main stem pair" (because it will modify the available model list), so vocals/instrumental in your case, and then you'll be able to select the models you want from the list
     
    • Like Like x 1
    • Interesting Interesting x 1
    • List
  3. Blu

    Blu Producer

    Joined:
    Jan 9, 2021
    Messages:
    199
    Likes Received:
    111
    The one you're using is currently the best combination of algorithms for stem separation AFAIK. I tried several algorithms and combinations and with this very same algorithm I had great success isolating drums, particularly kicks and snares ( hi-hats or cymbals in general become slightly muffled/blurred depending on circumstances) and good success with voice extracting, much better than not so long time ago but still not quite there in terms of artifacts. For drums only however, demucs 4 alone does a better job than this ensemble.
     
  4. curtified

    curtified Rock Star

    Joined:
    Feb 3, 2015
    Messages:
    669
    Likes Received:
    449
    you talking Demucs 4_ft? or do you have a pref?

    If I have a full song.. I usually do ensemble to make a inst first then o demucs 4_ft.. sometimes I do it once then take that rendered file and pass it back through the algo a few more times depending until it strips out artifacts.
     
  5. poly

    poly Platinum Record

    Joined:
    Sep 29, 2016
    Messages:
    285
    Likes Received:
    173
    Location:
    Hä?
    But what i didn't understand in Ensemble Mode is how i can set different models for the different stem pairs.
    If i set a model for a stem pair, switch to another stem pair, back to previous stem pair and the previous selected model here has gone.
    Or understand i something really wrong. How i understand it is that if i set UVR to Ensemble mode i can set different models to different Stem Pairs and then extract a song with this Ensemble in one action. :dunno:
     
  6. Blu

    Blu Producer

    Joined:
    Jan 9, 2021
    Messages:
    199
    Likes Received:
    111
    Yes
     
  7. jarredou

    jarredou Guest

    Emsemble mode is to process the same stem pairs with multiple different models "at the same time". Not different stem pairs, you have to do some manual clicking for that ! ;)
    But most of the models avaibable with UVR are only vocals or instrumentals models, so you can't use them to extract other stem pairs (bass, drums, etc...)

    Previous versions of UVR were able to run as a command line, it was usefull for batch processing. This is not possible at the moment with the new version.
     
    Last edited by a moderator: Mar 10, 2023
    • Interesting Interesting x 1
    • List
  8. curtified

    curtified Rock Star

    Joined:
    Feb 3, 2015
    Messages:
    669
    Likes Received:
    449
    Im going to attempt to keep updating this list if you all submit responses.Here is the ultimate list of Stem separation tools I have found:

     
  9. clone

    clone Audiosexual

    Joined:
    Feb 5, 2021
    Messages:
    7,388
    Likes Received:
    3,255
    Is there a quality rank order sometime on the horizon? I just want to know if any of this is worthwhile for a quality improvement vs ripX. "do it yourself and check" for me, means no you can't tell either. The other options are more involved, and I'd like to know a benefit vs effort opinion.
     
  10. mild pump milk

    mild pump milk Russian Milk Drunkard

    Joined:
    Mar 29, 2014
    Messages:
    2,786
    Likes Received:
    2,381
    Location:
    Russia
    Win 7 users? Works fine? Can't test it yet, unfortunately
     
  11. jarredou

    jarredou Guest

    The guy behind MVSEP made a quality checker based on SDR score. You can find the info about dataset and process used for evaluation here : https://mvsep.com/quality_checker/.

    The results are here : https://mvsep.com/quality_checker/leaderboard2.php
    You can find results for RipX, Fadr, GSEP, Splitter.ai, Moses.ai Lalal.ai, etc...
    (there's also a leaderboard with their "synth" dataset, but the "multisong" one is more a "real world" situation).

    unfortunately SDR evaluation doesn't tell the full story and has limitations (higher score is not always the best sounding) as explained here : https://source-separation.github.io/tutorial/basics/evaluation.html


    I'm using it with Win 7 without any problems ! ;)
     
    Last edited by a moderator: Mar 16, 2023
  12. jarredou

    jarredou Guest

    Nice ! It may be useful to also keep note of the algo/models they use (like Acon Acoustica was using Spleeter, they were the first to embed it in a DAW, RipX is (was?) using MDX models, maybe now they use Demucs4 as they do guitar/piano in the last version, lots of free websites are actually using Demucs with baseline models without any fine-tuning or retraining...)
     
  13. curtified

    curtified Rock Star

    Joined:
    Feb 3, 2015
    Messages:
    669
    Likes Received:
    449
    Yeah feel free to denote what service is using what algo. and ill add it to them. I think I know what ones use what. I know audiostrip went from Demucs3 to v4 recently.

    I would love to do ranking but I also think that's subjective per the user. My current personal fave is UVR for a broad one-click solution. It also does batch processing which is nice to set it and forget it.
     
  14. Rain Drum

    Rain Drum Noisemaker

    Joined:
    Jan 14, 2023
    Messages:
    27
    Likes Received:
    6
    Thanks for sharing and perhaps put it in your opening post, so more users will be able to find it right away :)
     
  15. jarredou

    jarredou Guest

    There are some ongoing and really promising academic research on music source separation with diffusion models. Current models are mask based models, they are just selecting parts of already existing audio very precisely. Diffusion models will regenerate audio, so it will be as "clean" as source, no more bleeding, no more frequency holes shifting around when other instruments are loud, etc... This should be a real big step forward, even if it will generate a new genre of artifacts (think about eyes and hand/fingers problem with midjourney ! :rofl:).
     
  16. curtified

    curtified Rock Star

    Joined:
    Feb 3, 2015
    Messages:
    669
    Likes Received:
    449
    you smart!
     
  17. curtified

    curtified Rock Star

    Joined:
    Feb 3, 2015
    Messages:
    669
    Likes Received:
    449
    so here for this. Any papers on it you could link us to? even a git hub?
     
  18. BobbyMonfrado

    BobbyMonfrado Member

    Joined:
    Nov 20, 2017
    Messages:
    26
    Likes Received:
    15
    There's a discord server for UVR and anything else audio seperation related here: https://discord.gg/vw2Bmfb5 (invite lasts 7days)

    Lots of chatter/help/info
     
    • Like Like x 1
    • Interesting Interesting x 1
    • List
  19. Menorah

    Menorah Producer

    Joined:
    Sep 21, 2013
    Messages:
    263
    Likes Received:
    92
  20. jarredou

    jarredou Guest

Loading...
Loading...