UVR5 the Best AI stem separation algo?

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

  1. omPro

    omPro Newbie

    Joined:
    Jul 7, 2023
    Messages:
    2
    Likes Received:
    0
    Actually you have not. In order to use Ensemble Mode one needs to use GPU Conversion, if this is not possible (maybe because the lack of an NVIDIA GPU) then it is not possible AT ALL to use Ensemble mode. All of you have failed to make this clear to those who are new to this.
     
    Last edited: Jul 8, 2023
  2. omPro

    omPro Newbie

    Joined:
    Jul 7, 2023
    Messages:
    2
    Likes Received:
    0
    And to add to that, modern Macs don't have NVIDIA graphics cards because Apple no longer supports them. So as long as no AMD graphics cards are supported, Mac users are screwed unless they use external solutions.
     
  3. Londoner

    Londoner Member

    Joined:
    Oct 30, 2021
    Messages:
    53
    Likes Received:
    19
    Which model is MDX23 v2/2.1 from this pull-down menu in UVR5 settings?

    Screenshot 2023-07-08 at 23.01.39.png

    Or does this need to be loaded separately? Thanks!
     
  4. livemouse

    livemouse Producer

    Joined:
    Feb 13, 2016
    Messages:
    289
    Likes Received:
    137
    Tried to install this using Linux via wine but I guess it needs CUDA or something. Nevertheless, all it did was crash.

    However, The website service Fadr seems to work pretty well. Can rip a few tracks a day with a free account.

    Does anybody do this stuff on Linux with a non-Nvidia video card? Looking for tips.
     
  5. Ryck

    Ryck Guest

  6. jarredou

    jarredou Guest

  7. jarredou

    jarredou Guest

    You could have asked how to do when you have no GPU, instead of that rant, and I would have explained how to do... But no. :dunno:
     
    Last edited by a moderator: Jul 9, 2023
  8. GeekedGlitch

    GeekedGlitch Ultrasonic

    Joined:
    Jun 13, 2023
    Messages:
    94
    Likes Received:
    25
    @jarredou Hey, I've been trying to use your amazing product. I want to say thanks for sharing this instrument, I feel that it's priceless.

    Can you please share with me a quick way of downloading processed audio? I've never used collab; as I understand, it's best to zip all tracks inside collab, to somehow speed up, since Collab doesen't give too much kbps... What code should I use to quickly zip my output folder and then download it?
     
  9. jarredou

    jarredou Guest


    Best way is to save the processed audio directly on your mounted Google Drive, and then to download the files from Google Drive, it will be way faster than downloading them from Colab (and if you download the folder containing all the processed file, it will be automatically zipped before the download start).
     
  10. herznote

    herznote Ultrasonic

    Joined:
    Jul 11, 2020
    Messages:
    57
    Likes Received:
    32
    need help ! ;) what are the best settings to separate a vocal from an intro interlude ? thanks in advance
     
  11. GeekedGlitch

    GeekedGlitch Ultrasonic

    Joined:
    Jun 13, 2023
    Messages:
    94
    Likes Received:
    25
    I think Kim Vocal 2 is the best currently available model for vocals
     
  12. Martel

    Martel Platinum Record

    Joined:
    Jan 8, 2023
    Messages:
    398
    Likes Received:
    176
  13. herznote

    herznote Ultrasonic

    Joined:
    Jul 11, 2020
    Messages:
    57
    Likes Received:
    32
    where can i get it? didnt see it in uvr5
     
  14. Martel

    Martel Platinum Record

    Joined:
    Jan 8, 2023
    Messages:
    398
    Likes Received:
    176
    @jarredou , I see that there's a new sheriff in town.

    Can we use those in UVR5 ? MDX23C, Grow 128, Hop 1024

    If not, where can we use those ? Are they available yet ?

    Thanks for your time, always appreciated.
     
    Last edited: Jul 16, 2023
  15. Martel

    Martel Platinum Record

    Joined:
    Jan 8, 2023
    Messages:
    398
    Likes Received:
    176
    You can use MDX-Net: 292, 496, 406, 427, Kim Vocal 1, Kim Inst + Demucs ft


    In UVR5 it's under: Ensemble - Vocal / Instrumental
    MDX-Net: Kim Inst
    MDX-Net: Kim Vocal 1
    MDX-Net: UVR-MDX-NET Inst Main
    MDX-Net: UVR-MDX-NET Inst_full_292
    MDX-Net: UVR-MDX-NET_Main_406
    MDX-Net: UVR-MDX-NET_Main_427
    Demucs: v4 | htdemucs_ft



    MDX-Net info:
    Batch Mode

    Demucs info:
    CLI version
    Model: hdemucs_ft
    Default options

    Time Elapsed: N/A

    Algorithm: Average / Average
     
    Last edited: Jul 16, 2023
  16. GeekedGlitch

    GeekedGlitch Ultrasonic

    Joined:
    Jun 13, 2023
    Messages:
    94
    Likes Received:
    25
    @jarredou I got one more question. Do overlapping and memory chunk size affect quality of the demixed audio? As if it were able to create artifacts and/or bleeding
     
  17. jarredou

    jarredou Guest

    These new models are only available on mvsep1.ru (who trained them) at the moment. The new ensemble is premium only.
    UVR will publish new model with the upcoming update (currently in beta).


    The mutlisong leaderboard is more complete and is based on real songs (the other is synthetic, mixing unrelated stems together to create a fake eval dataset) : https://mvsep.com/quality_checker/leaderboard2.php?sort=vocals


    They do, chunk_size should be set to the maximum you can with the available memory ressources (song length can also influence memory usage), it's quite already at the mamimum in my MDX23-Colab default settings (for Colab free accounts).
    Overlap is also improving quality, but has less influence than chunk_size. It aslo help reducing some artifacts happening at audio chunks boundaries.
    Overlap=0.8 is good value for balance between quality and bearable processing time.
     
    Last edited by a moderator: Jul 17, 2023
  18. Martel

    Martel Platinum Record

    Joined:
    Jan 8, 2023
    Messages:
    398
    Likes Received:
    176
    I have 588 credit still on mvsep1.ru.

    I'll give it a shot.

    If I understand correctly, the second Ensemble called ''MDX23C'' Ensemble is the one that is currently giving the best result for vocals. Do I understand correctly?

    Thanks, I appreciate your help.
     
  19. jarredou

    jarredou Guest

    That's it !
     
  20. SomeOtherGuy

    SomeOtherGuy Member

    Joined:
    Sep 4, 2014
    Messages:
    29
    Likes Received:
    16
    Location:
    Thailand
    Is there any model which can remove thousands of screaming fans from a live recording?
     
Loading...
Loading...