That's why you should never blindly trust LLMs (Musk's one specially): OpenVINO™ AI Plugins for Audacity

 

Is that really same model in Logic? That Logic sample is worse than mvsep, you can hear those phasing issues..

 

Bs roformer sw is logic pro but

1.Mvsep is ensemble of different drum models, including logic one(sometimes better sometimes worse)
2.by default settings mvsep use an instrumental model first and remove vocal, and then use drum models... But logic extract drum from the whole song...

 

can i have the original file?

 

Is becruily better than UVR MDX23C-Instvoc HQ?

 

yes

 

Where is the best site/app to run becruily?

 

Spoiler: Source: Perplexity KI

The best site or app to run Becruily models, such as "mel-band-roformer-instrumental," is through Ultimate Vocal Remover (UVR), specifically UVR version 5 or higher, as users report successfully running Becruily models within this software8. UVR is open-source, free, and can be downloaded and run on most major operating systems35. There is also a strong community of users and regular updates, making it a reliable platform for Becruily models.

If you prefer browser-based processing or want to avoid local installation, you may also find UVR-related online or Google Colab notebooks referenced in the community710. These can allow you to run Becruily models directly in your browser, though performance and model support may vary compared to a local installation.

In summary: The recommended app is Ultimate Vocal Remover (UVR 5.0 or later), available as a download or in Colab notebook form. This has the broadest compatibility and support for Becruily models358710.

 

Mvsep has it

 

Not all demux have the same efficacy. Some sources have different situations. Spectralayers 12 is better (IMO) to separate Vocals, Drums & Basses. Inclusive you can split different kinds of drums (Kicks, Snares, Hi-Hats, Toms & Cymbals in a second split) with great results. Other sources, like pianos, synths and guitars are very rare to split, and some times it shows parts of guitars on keyboards splits. You can copy those ghosts sounds and paste it on other layers and get a better track. (SpectraLayers is specially good doing that kind of manually processes), among you can select some indeseable freqs or noises and delete it, like if you are working on Adobe's Photoshop. Overall, it's a long and tedious work but it's possible to get great results. It's trial and error.

 

This is a good example of a song with a drum loop that I would like to extract (previous examples of this track were MP3 files). This is also the quality that I would like to keep all of my audio at. UVR and mvsep only have 44.1 and would need to be upsampled to 48.

WAV File (24/48k):

 

I don't understand what I'm looking at. How does this list show any comparison? It lists six other separation tools.

Comparison with Alternatives
Here’s how UVR compares to other leading vocal and drum separation tools in 2025, based on available data:

1. LALAL.AI:
   • [...]
2. Moises.ai:
   • [...]
3. iZotope RX 11:
   • [...]
4. SpectraLayers 11:
   • [...]
5. MVSEP:
   • [...]
6. RipX:
   • [...]

 

Last edited: Aug 5, 2025

How is
OpenVINO™ AI Plugins for Audacity???

 

MDX-NET Inst HQ5 is often suggested, but on UVR I can only find HQ4. Is there an HQ5 and if so where is it?

Another issue are the names of models. For example, many users recommend MB Roformer v1 or mel-band-roformer-instrumental (recommended by PulseWave), but there is no such models on UVR. There are, however, a bunch of similar named models. How can a user figure out which one is recommended?

This app is a mess. I'm sure good results are to be found somewhere in there, but come on . . . is this intentionally designed to be a labyrinth?

 

Last edited: Aug 5, 2025

SpectraLayers 11 isn't good as SpectraLayers 12

 

I would like to thank everyone who contributed to this thread. There is no clear answer to my question at this time. Many users seem to get good results from UVR (via various models and combinations/ensembles) as well as other apps and online services. In my case, I need 24bit/48k quality. That eliminates many contenders, including UVR. Until something better comes along, I'll use Logic Pro in combination with ________ (not sure what). I deleted UVR from my hard drive . . . too fiddly and not high enough fidelity.

Rock on!!!

 

Hi tommyzai, I've been following this thread the past few days to learn things from users like @akbarz and because I find it interesting. I'm just wondering why 24bit/48kHz is necessary for the task at hand. I'm sure you have a reason, and I'm not trying to be annoying here, but reading this thread had me wondering why the best possible quality at the bleeding edge of this new technology isn't adequate for the task of teaching kids the fundamentals of music making. The lesson should be equally valuable whether the stems are 44 or 48kHz. You don't need to add them to your personal music archive!

 

Free alternatives: Audacity www.audacityteam.org

Music separation: www.audacityteam.org/download/openvino
Separate a mono or stereo track into individual stems -- Drums, Bass, Vocals, & Other Instruments.

Releasing Spleeter: Deezer Research source separation engine
https://deezer.io/releasing-spleeter-deezer-r-d-source-separation-engine-2b88985e797e

How to Separate Vocals from Instrumentals with Audacity's New AI Plugin

Get the plugin here and follow along with the installation and setup process.
https://github.com/intel/openvino-plugins-ai-audacity/releases

Spleeter.Online - https://spleeter.online

 

Last edited: Aug 6, 2025

I appreciate the respectful manner by which you are asking this question . . . and it's a good question. Yes, the primary reason for my music production is education, but I/we have everything set up in 24/48 (all recording, audio samples, project settings, latency compensation, etc., and it would be easier and better to keep it all at that format. If possible, I'd like to avoid a conversion with downgraded quality, and I certainly don't want to lower quality and change workflow to suit a stem sep that is using an old industry standard.

If 44.1 was the only issue with UVR, I would use it. I might revisit this thread and UVR down the road, but at present . . . instead of helping with music creation and education, UVR is confusing, frustrating, and counterproductive. Using it for stem separation is a HUGE rabbit hole. Some . . . many enjoy experimenting and investigating models and parameters and are fascinated by it all. I get that. But in my situation . . . I'm desperately trying to keep things simple, straightforward, and powerful, but that's not easy nowadays.

 

Last edited: Aug 6, 2025

Any thoughts on Moises.ai?
The paid pro account will yield 24 bit/48k.