UVR5 the Best AI stem separation algo?

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

  1. curtified

    curtified Rock Star

    Joined:
    Feb 3, 2015
    Messages:
    669
    Likes Received:
    451
    I noticed this was only mentioned once before in this forum, so I thought I'd give it its own post. I've tried several stem separation algorithms like izotope RX, Acon Acoustica, lalal.ai, audiostrip, and even some freeware. However, this particular algorithm has been the best one I've found so far. What sets it apart is how it combines multiple algorithms to create a higher-quality final product. You can even hear reverb tails in vocal stems and, if you're lucky, you can extract individual drum sounds from certain materials. Setting it up might be a bit complicated, but the end result is definitely worth it. Im excited for whats to come in 2023 the year of the AI boom!

    UVR AI STEM SEPARATOR
    Download here:
    https://github.com/anjok07/ultimatevocalremovergui/releases

    What’s the best settings?
    https://github.com/Anjok07/ultimatevocalremovergui/issues/344

    Im currently using this mode:
    MDX-Net: UVR-MDX-NET Inst 3 + UVR-MDX-NET Inst Main + Kim_vocal_1 + Demucs: v4 | htdemucs_ft - Shifts: 2 - Ensemble Algorithm: Avg/Avg

    Settings leaderboard:
    https://mvsep.com/quality_checker/leaderboard.php

    MORE READING HERE:
    https://www.reddit.com/r/IsolatedTracks/comments/vuavwq/ultimate_vocal_remover/

    Also adding this list of all the current stem separation algos out there:

    UPDATE!
    I added many of the FREE UVR algos to my FREE Voice cloning app:

    https://www.tryreplay.io/ It runs locally! No server is needed! We even have stem separation, de-reverb built in. We're about to launch a plugin in a few weeks too!

    We've also built an AI model hosting platform over at https://www.weights.gg/ (with 75k+ users & 16k+ models). Think of it as a Splice for AI models. Please download or contribute your models to our library!

    We're open to feedback from the music-making community, so please share any thoughts here or on our Discord(s).

    Please bring any questions to this thread - https://audiosex.pro/threads/free-a...ning-app-with-built-in-stem-seperation.72525/
     
    Last edited: Oct 26, 2023
    • Like x 12
    • Agree x 7
    • Interesting x 6
    • Winner x 2
    • Love it! x 1
    • List
  2.  
  3. clone

    clone Audiosexual

    Joined:
    Feb 5, 2021
    Messages:
    7,440
    Likes Received:
    3,280
    Have you done a test comparing the results of Hit 'n Mix RipX with this?
     
  4. curtified

    curtified Rock Star

    Joined:
    Feb 3, 2015
    Messages:
    669
    Likes Received:
    451
    I have not (yet)
     
  5. Legotron

    Legotron Audiosexual

    Joined:
    Apr 24, 2017
    Messages:
    2,181
    Likes Received:
    2,113
    Location:
    Hyperborea
    This is actually really good, best at the moment IMO
     
    • Like Like x 2
    • Agree Agree x 2
    • List
  6. jasper700

    jasper700 Newbie

    Joined:
    Mar 8, 2021
    Messages:
    1
    Likes Received:
    1
    There's a new leaderboard for the 4 and 6 stem algos here:
    https://mvsep.com/quality_checker/leaderboard2.php?sort=vocals

    Note that you can sort the list depending on the stem you're most interested in. They don't do as well as the best ensembles - or even the best single algos - at Vocal/instrument splitting. But they DO give you 4 (or 6) stems. Looks like Demucs: v4|htdemucs_ft is still top dog.

    I'd love to be able to import stems from that algo into RipX, because the RipX software is so cool.
     
  7. Martel

    Martel Platinum Record

    Joined:
    Jan 8, 2023
    Messages:
    398
    Likes Received:
    176
    It is not the best quality at the moment but it is definitely the best in its price range as it is free hahaha.

    Joke aside, there's Audioshake that does a better job at stems splitting but at 12$ USD per wav stems. So a full single track is like 60$ USD.

    They also reserve themselves the right to keep your money and not let you download the song you splitted if they discover that you are using a commercially released song and that you don't have the rights to it.

    So yeah, UVR5 definitely the best accessible but just not the best quality.

    Curtified, those are the best settings for Vocals and Instrumental splitting in UVR5.

    For the rest, (ie Drums alone or Bass alone or Others alone) Htdemucs_ft alone is giving the best results.

    Great informative post by the way.

    i discovered it because of a user on the ister sites after trying the Demucs platform on Google colab.

    The dev is definitely worth a coffee.
     
    • Agree Agree x 2
    • Interesting Interesting x 1
    • List
  8. clone

    clone Audiosexual

    Joined:
    Feb 5, 2021
    Messages:
    7,440
    Likes Received:
    3,280
    That should be called Audio Shakedown.
     
    • Funny Funny x 6
    • Agree Agree x 3
    • Like Like x 2
    • List
  9. BagelShaggerz

    BagelShaggerz Platinum Record

    Joined:
    Apr 19, 2022
    Messages:
    475
    Likes Received:
    182
    This looks very interesting thanks, but how do you download GitHub links? It looks like a ton of files and I don't know what I'm doing.

    Nvm, I got the setup file from the other link. But when I'm on the page with the loads of files with the green download link drop down button page I usually get linked to looks confusing and I never know how to download from there.
     
  10. Lois Lane

    Lois Lane Audiosexual

    Joined:
    Jan 16, 2019
    Messages:
    4,786
    Likes Received:
    4,703
    Location:
    Somewhere Over The Rainbow
    Whenever trying to download anything from GitHub it is as if I'm a blind dyslexic and sometimes just bow my head in despair, turn sheepishly on my idiot heels in shame and walk away.
     
  11. BagelShaggerz

    BagelShaggerz Platinum Record

    Joined:
    Apr 19, 2022
    Messages:
    475
    Likes Received:
    182
    LOL That's exactly what it's like sometimes. If I get directed to a page that doesn't just have just one link to download but instead to a page with the entire listing of files, I start sweating bullets like I'm reading a foreign language in a new country cause I took the wrong flight.

    That's why Home Alone had Kevin accidentally take a flight to New York instead of Bali. Cause we know how that would end, Kevin would start asking people how to get back to Chicago but no one where would know what the fuck he's talking about. Then by the end of the week he'd have given up and accidentally smoked a joint and done a line yay with Bali enforcement to take the load off. But...

    Then like you know, would have been put to death by squad fire. The credit would roll in 14 minutes and no one would have ever watched Home Alone 2.

    [​IMG]
     
  12. ArticStorm

    ArticStorm Moderator Staff Member

    Joined:
    Jun 7, 2011
    Messages:
    7,731
    Likes Received:
    3,945
    Location:
    AudioSexPro
    DEMUCS3 modelB is quite decent.

    Trying DEMUCS4 HT now with High Quality. Lets see how it compares.
     
    Last edited: Feb 28, 2023
    • Like Like x 1
    • Interesting Interesting x 1
    • List
  13. curtified

    curtified Rock Star

    Joined:
    Feb 3, 2015
    Messages:
    669
    Likes Received:
    451
    Wait you can get more stems out of it than the 4 bass, drums, vocals, other? please share!
     
  14. jarredou

    jarredou Guest

    Yes, there's a 6stems version of Demucs4ht (+ guitar & piano, even if the piano model is not really good).

    GSEP is also one of the best even if its SDR scores are lower than others, it sounds really good to human ears (and still free for the moment, you only need to create an account !)
    It also has guitar & piano models
    https://studio.gaudiolab.io/
     
    Last edited by a moderator: Mar 1, 2023
    • Interesting Interesting x 2
    • Like Like x 1
    • List
  15. jarredou

    jarredou Guest

    There's also https://x-minus.pro/, not tried but it uses MDX/Demucs models, and they also have a premium thing with their own trained models (piano, guitar, horns...).
     
  16. curtified

    curtified Rock Star

    Joined:
    Feb 3, 2015
    Messages:
    669
    Likes Received:
    451
    Youre the goat for suggesting gaudiolab. Thank you!
     
  17. aleksalt

    aleksalt Producer

    Joined:
    Jul 1, 2013
    Messages:
    383
    Likes Received:
    105
    After reading this thread I recollected one unit (it's 9 years old) and what people say about it, mostly 2 phrases:
    "This is fantastic" or "What is this for?"

    It's about PITCHMAP by Zynaptiq

    from the official description:

    "PITCHMAP is the world's first and only plug-in that allows processing the pitch of individual sounds within mixed signals in real-time. It lets you change the melodies and harmonies of any recording by simply playing your own on a MIDI keyboard or creating a pitch map using our stream-lined GUI, and optionally corrects any tuning issues at the same time"

    So, according that, having a recording of any orchestral theme (Bach, Mozart, or name what you'd prefer)
    you can change it to unrecognizable theme with new chord progressions and melodies...

    Next, let's combine those plugins above with Pitchmap, take a top notch recording and 1st extract stems and 2nd: change the pitch of notes and places, where you desire, for example: violins, violas. cellos, basses and then join them together again.
    Is that real?
     
  18. ArticStorm

    ArticStorm Moderator Staff Member

    Joined:
    Jun 7, 2011
    Messages:
    7,731
    Likes Received:
    3,945
    Location:
    AudioSexPro
    v4 HT HQ mode improves a lot compared to v3 if you throw in the same song.
     
  19. curtified

    curtified Rock Star

    Joined:
    Feb 3, 2015
    Messages:
    669
    Likes Received:
    451
    Pitchmap and zplane retune are both amazing tools with different algorithms, and although they are similar, I find myself preferring zplane retune more frequently than Pitchmap. With these tools, it is possible to change the pitch of notes within samples in real time, similar to a real-time melodyne effect. However, it is worth noting that while these tools can go unnoticed in some cases, they can also leave artifacts. Nevertheless, if used correctly and masked within other elements of a song, these artifacts can be concealed. Overall, we are fortunate to be living in an era of incredible technological advancements.
     
  20. curtified

    curtified Rock Star

    Joined:
    Feb 3, 2015
    Messages:
    669
    Likes Received:
    451
    have you tried ensemble mode?
     
  21. ArticStorm

    ArticStorm Moderator Staff Member

    Joined:
    Jun 7, 2011
    Messages:
    7,731
    Likes Received:
    3,945
    Location:
    AudioSexPro
    i am using a cloud service, which offers only a few algos, but i dont have to run it on my computer.

    there is no ensemble mode
     
Loading...
Loading...