HI all !
Can someone share their best experience/settings for Ultimate Vocal Remover's Ensembles Mode ??? Vocal/Instrumental pair.
Thanks

 

If GPU=CUDA only, then yeah. Lots of AI stuff works on Apple Silicon GPU, often things start out as CUDA only and then general GPU support comes later. That said, the GPUs are a definite weak link for the M1+ series so far. I don't know how much of that is hardware power and how much of it is a lack of available adaptable codebase to port things optimally to apple GPUs. But the end result is the same - we all fantasize about buying a PC rig just for a juiced up Nvidia GPU.

 

I forgot to mention here that ByteDance SAMI algo can now be used through their new Ripple iOS app (only available in US atm, but that can be tricked changing your apple ID localization... current version has a chunk artifacts problem every 8 seconds. Their API is performing better than Ripple, but is only accessible with a Chinese ID, it seems...)
https://apps.apple.com/us/app/ripple-music-creation-tool/id6447522624?uo=4

Their academic paper will be published on arXiv in the next days.

 

Last edited by a moderator: Sep 6, 2023

SDR?

 

https://mvsep.com/quality_checker/multisong_leaderboard?sort=vocals

 

great, cant wait to see public model that achieves sdr 20 on instrumental

 

Hello! I'm new here. Help me please. What algorythm will be better to get most accurate piano stem from audio file? Thanks in advance!

 

https://studio.gaudiolab.io/gsep
https://mvsep.com

are the both best available - but far from perfect -. Demucs4 has also a piano model in its "6s" version but it performs lower than these 2 ones.

 

I tested the performance of the Stems 2.0 algorithm of Virtual DJ 2023 some time ago:

Nvidia RTX 3070 (Windows) - 30.3 times as fast as real time - over 200W power consumption
M1 GPU MacBook Air 2020 - 12.1 times as fast as real time - 15W power consumption

This is quite impressive. If you calculate the power consumption over time, the M1 GPU needs only 18.8% of the power the RTX 3070 needs for the same task.

Sure, the RTX is almost three times as fast, but the M1 in the Air 2020 has the smallest 7 core GPU of all Apple Silicon chips. So the bigger M1 or M2 chips with more GPU cores should be quite capable for AI calculations and use way less power.

 

Thank you so much for answer!

 

Worth noting to Mac users: starting with Big Sur, you can run iOS apps inside MacOS on Apple Silicon machines (not Intel).
Would love to hear some samples of SAMI-Bytedance algorithm via Ripple iOS app if anyone has a capable machine and wants to throw an old standard song at it!

 

Last edited: Sep 17, 2023

To newbies, I always recommend Lalal.ai - its a website, with its own algo, super user friendly and better than all the open source stuff I've tried. It's not free but you get some trial minutes, which should be enough for one stem.

 

The dev needs to allow it, and this dev does not. Just tried. Might check it out on my phone.

 

Thanks for suggestion! In my case Gaudio Studio made it more better and clearly. I use LalalAI sometimes. It was my file for extraction piano and LalalAi had more besides artifacts

 

Lalal.ai is outclassed even by Demucs 3.0 mdx_extra

 

phew the leaderboard is confusing as hell. So what is izotopes algorithm called, is it on the leaderboard?

MVSEP is at the top of the list, so i might shootout that vs izotope, but last time i did this with spleeter it was a big waste of time, spleeter was the same or worse vs izotope except it took ages of setup and command line nerd-nirvana

 

Bro, SOOO much has changed since Izotope/spleeter was the best algorithm. UVR, RipX are the best software (with a GUI) to utilize the newer algorithms, both are incredible pieces of software. There are so many different forks or implementations of different algorithms, but Demucs, and (MDX2C) MDX2.3 for general separation tasks are in a whole other universe compared to were the spleeter algorithm left off. In short you're in for a pleasant surprise.

Read this to get caught up, or take a deep dive:
https://docs.google.com/document/d/17fjNvJzj8ZGSer7c7OFe_CNfUKbAxEh_OBv94ZdRG5c/edit#heading=h.roiuj54hzww3

 

Last edited: Sep 27, 2023

Let me spare you the wasted time,
There is no commercial software that can compete with Demucs 4.0, let alone MVSEP.
And SAMI-Bytedance may be the absolute best but we dont have any actual audio output stem samples of it to analyze and verify that it's actually a hifi audiophile passable stem model.

Some of the past leaderboard scores can be criticized due to the actual output stems being full of junk/mis-assigned frequencies, bleed, artifacts, spectral holes, etc. Furthermore any model trained on Lossy is automatically disqualified from an Audiophile standpoint. Some of these models have terrible bleed but score ridiculously high, so it is possible that the Scoring algorithm is flawed, from an audiophile standpoint.

Separation Accuracy, Stem Cleanliness, and non-destructiveness should be the true metrics.
Some of these models that generate crap stems full of garbage frequencies have scored ridiculously high.

If you want to do quick Remixes, something like Izotope is perfectly capable and "good enough" probably for most people.
If youre seeking the best quality AI stem model for audiophile/hifi purposes, you wont find it on commercial software.

 

RipX is very impressive. Way better than Izotope Music Rebalance with what I have used them on. But it is rather slow processing.

Acon Digital just released a plugin called Remix which I have not read much about or tried at all. But I am very interested to soon, because their general audio restoration plugins are very good. They describe it as "Real-Time" and low cpu usage. But these companies make lots of claims, as we all know.

 

Hi,

can somebody explain to me, what is meant by the UVR-algorithms, which have the word "MGM" in it? Is it made for MGM (Metro-Goldwyn-Mayer) Recordings?