UVR5 the Best AI stem separation algo?

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

  1. Dyslexicon

    Dyslexicon Noisemaker

    Joined:
    Mar 19, 2023
    Messages:
    23
    Likes Received:
    4
    @jarredou could you DM me please, have a few ideas to discuss thanks
     
  2. Abner

    Abner Newbie

    Joined:
    Oct 26, 2020
    Messages:
    12
    Likes Received:
    1
    Is there an algo out there able to separate lead vocals from background vocals? And what about guitar solos from rhythm guitars?
     
  3. Menorah

    Menorah Producer

    Joined:
    Sep 21, 2013
    Messages:
    263
    Likes Received:
    92
  4. StormChaser

    StormChaser Producer

    Joined:
    Jan 16, 2021
    Messages:
    290
    Likes Received:
    115
    Have you tried StemRoller? For a free bit of software it does a very good job of Stem separation.

    It also allows you to search YouTube from with the the app and pull down and separate the track you are looking for.

    https://www.stemroller.com/
     
  5. kevdel10

    kevdel10 Noisemaker

    Joined:
    May 14, 2022
    Messages:
    4
    Likes Received:
    4
    Ah, I see! Thanks for the thorough explanation.
     
  6. deton24

    deton24 Member

    Joined:
    Jun 23, 2023
    Messages:
    9
    Likes Received:
    14
    Check the model list here:
    https://docs.google.com/document/d/17fjNvJzj8ZGSer7c7OFe_CNfUKbAxEh_OBv94ZdRG5c/edit?usp=drivesdk
    I'm trying to keep it updated with the algos, Colabs, tips and other stuff.

    BTW. Jarredou released V2 MDX23 Colab with newer models (I think he fixed the issue with vocals stem mentioned above), and a day later, slightly better than Kim vocal 2:

    Vocal Ft model was also released in UVR 5 (and Audio Hacker (a.k.a. HV) Colab)
    https://colab.research.google.com/github/NaJeongMo/Colab-for-MDX_B/blob/main/MDX-Net_Colab.ipynb

    Now MDX23 can work with 8GB VRAM decent Nvidia GPUs. I described some workarounds for 6GB VRAM cards as well.

    Also check new HQ3 fullband instrumental model in UVR5 download center.
    All links are in the doc above.

    Also Drumsep Colab with fixed model file for separating single drum instruments.

    And the best ensemble for now in UVR 5 GUI is:
    (#3842)
    Kim Vocal 1 + 2 + Kim Inst + HQ3 + Voc FT + htdemucs_ft | Avg/Avg | Chunks: O
    N
     
    Last edited: Jun 24, 2023
  7. jarredou

    jarredou Guest

    Nice to see you here Deton24 ! :)
     
  8. maskie

    maskie Noisemaker

    Joined:
    Aug 20, 2016
    Messages:
    6
    Likes Received:
    3
    I downloaded drumsep from the above github page and put it in UVR5 models .Changed the name and check sum and it splits drums into 4 stems,kick,snare .cymbals,toms.Seems to work well but in order to use it as a separate model in UVR5,it seems i have to edit the py script and .json.Any body who knows more about this than I ,please help. Thanks!
     
  9. AudioSep

    AudioSep Newbie

    Joined:
    Jul 1, 2023
    Messages:
    1
    Likes Received:
    0
    Hello Maskie, I'd like to know how to add drumsep to UVR5. How did you go about it as I use it through a notebook but it's manual and tedious and having it run within UVR would be great.
     
  10. deton24

    deton24 Member

    Joined:
    Jun 23, 2023
    Messages:
    9
    Likes Received:
    14
    Go to UVR settings and open application directory.

    Find the folder "models" and go to "demucs models" then "v3_v4".

    Copy and paste both the .th and .yaml files, and it's good to go.
     
  11. HansPaetsch

    HansPaetsch Newbie

    Joined:
    Jul 2, 2023
    Messages:
    14
    Likes Received:
    0
    Hello,

    I have just come across the software and am very impressed with its range of features. What I am missing, however, is some kind of instruction manual. Would someone please explain to me which model and which processing method is used for what?

    In the first case, I have a rare TV recording from the 80s (
    that I would like to take apart into its component parts.

    In the second case, it is a 78rpm digitised record: https://ia801603.us.archive.org/27/items/78_chinaman_lilian-harvey-willy-fritsch-schroder-beckmann_gbia3002184a/CHINAMAN - LILIAN HARVEY - WILLY FRITSCH.flac (i will do a cleanup with rx9) that I would also like to take apart.

    If anyone could help me with this, I would be very grateful.

    Many greetings from Germany

    HansPaetsch
     
    Last edited: Jul 2, 2023
  12. iluvhiphop

    iluvhiphop Guest

    Honestly, I can't really tell them apart. The music I seperate with these AI tools is usually recordedly really poorly. Think Afghan 70s psychedelic polka. These models are usually trained on more modern music, so the quality of seperation is kind of hit and miss when you seperate other things. I try different algorithms and pick whatever is best, and it's almost never whatever is topping on the leaderboards.
     
  13. Legotron

    Legotron Audiosexual

    Joined:
    Apr 24, 2017
    Messages:
    2,181
    Likes Received:
    2,113
    Location:
    Hyperborea
    @HansPaetsch Demucs 4 - htdemucs_ft for separating drums, bass, vocal, other
    Demucs 4 - htdemucs_6s for drums, bass, vocal, guitar, piano, other
     
  14. Legotron

    Legotron Audiosexual

    Joined:
    Apr 24, 2017
    Messages:
    2,181
    Likes Received:
    2,113
    Location:
    Hyperborea
    @deton24 @jarredou How do you force UVR to use separate GPU? I just noticed my RTX 3050 isn´t doing anything, even if GPU conversion is ticked from the GUI.
    Could there be some conflict, because I have intergrated GPU, which I use for display? I mainly bought the separate GPU for GPU Audio and AI stuff, since I don´t play games
     
    Last edited: Jul 2, 2023
  15. HansPaetsch

    HansPaetsch Newbie

    Joined:
    Jul 2, 2023
    Messages:
    14
    Likes Received:
    0
    @Legotron. Thank you for your reply. And which mode? Ensemble.
     
  16. Legotron

    Legotron Audiosexual

    Joined:
    Apr 24, 2017
    Messages:
    2,181
    Likes Received:
    2,113
    Location:
    Hyperborea
    @HansPaetsch I just use the htdemucs_ft and htdemucs_6s, no ensembles. From advanced demucs option, i set shift=1, overlap=0.99, chunks=Auto and only mixer mode ticked. It´s definetely not perfect, but pretty damn good, of course depending on your use case.
     
  17. maskie

    maskie Noisemaker

    Joined:
    Aug 20, 2016
    Messages:
    6
    Likes Received:
    3
    @deton24 thanks ! I just did a quick hack,yours works great.
     
  18. Dyslexicon

    Dyslexicon Noisemaker

    Joined:
    Mar 19, 2023
    Messages:
    23
    Likes Received:
    4
    I didnt see a .yaml file in the repo where the modelo.th file came from.
    Does the .yaml file have to be manually created? No idea how to create one.
     
  19. Legotron

    Legotron Audiosexual

    Joined:
    Apr 24, 2017
    Messages:
    2,181
    Likes Received:
    2,113
    Location:
    Hyperborea
    I just copied those 2 files in demucs folder and it showed up as demucs model, no editing
     
  20. deton24

    deton24 Member

    Joined:
    Jun 23, 2023
    Messages:
    9
    Likes Received:
    14
    https://docs.google.com/document/d/...bAxEh_OBv94ZdRG5c/edit#heading=h.rz0d5zk9ms4w

    MDX23 v2/2.1 Colab is currently better than htdemucs_ft
     
Loading...
Loading...