__How to remove artefacts from an inverted acapella?_____
This section is old, and “cleaning inverts” in current models section can provide more up-to date solutions.
0) Currently, GSEP is said to be the best in cleaning inverts. But at least for vocal you can use some MDX model like Kim, or even better MDX23 from MVSEP beta.
- by charm
(rather outdated) Use Vocal_HP_4BAND_3090_arch-124m.pth at 0.5 aggressiveness, tta enabled
then use any model u like with Vocal_HP_4BAND_3090_arch-124m.pth instrumental results to filter out any vocals that weren't detected as vocals with Vocal_HP_4BAND_3090_arch-124m.pth model
combine the two results
then use model ensemble with whatever models u like (i used HP2 4BAND 1 and 2)
drag both vocal hp 4band+another model and ensemble results into audacity
use the amplify effect on both tracks and set it to -6.03
render
then use StackedMGM_MM_v4_1band_arch-default.pth
tbh vocal models even at 0 aggressiveness really help inverts
Or 0.3
I mostly use acapellas for mashups and remixes, so the little bit of bleed i get at 0.0 aggressiveness is fine
drums-4BAND-3090_4band.pth
0.5 optionally (less metallic sound)
2)
Utagoe (English version with guide and error messages translated by Anjok) - if the invert isn't good, then try it, but it’s not the best (although rather better then Align Inputs in UVR).
(Utagoe4 by Jake - explained later at the bottom;
Frasier also created Utagoe 4 - reverse engineered fork with new tricks vs the v3, and later also with intention to outperform the Jake's one, consider comparing them)
Settings for Utagoe (OG) by HV:
“if your tracks don't invert perfectly” (even when aligned)
“if it's perfectly inverting”:
Some other settings:
“It has a weird issue sometimes tho, even when everything is perfectly aligned and inverts perfectly, utagoe misses some places, and it won’t insert for a second or so”
by Mixmasher00
“There is no actually settings depending on songs, but that is what I use, which is the default one.
Going higher than 1.3 [of extractable level] imo won't do good at cleaning. Additional tip too, if you want to do just an "invert" and "keep the original vocal volume" just choose "by waveform”.
I have been using Utagoe for inversions recently because it keeps the original volume of the vocals, and then I ran it on UVR or MDX. If the chunks are soft, I prefer using UVR but if there are chunks [that are] heavy like drums, I'd use MDX.
Also, I find it better to clean an invert via UVR or MDX than Utagoe because it's better and cleaner without destroying the vocals“
- “When using Utagoe, or UVR5, for aligning inputs, and inverting them, I get this really strange cracking noise, that not even doing a vocal separation with AI later can get rid of. [Does anyone] know what could possibly be causing this?
So, I actually found a solution to this, for anyone running into a similar issue.
With inversions like this, you're already gonna have to use AI to get rid of the leftover noise, since it's not gonna be a perfect inversion. So, the solution isn't to get a perfect one, it's simply to get rid of the noise that the AI cannot recognize, right?
With this particular inversion, I originally was using really compressed MP3s from around 2007 for the instrumentals, because the lossless versions of the instrumentals were lost media, up until a few months ago.
I thought it was odd, because I don't remember this noise being an issue with the MP3s, and that's when it hit me, MP3s cut off the noise, with compression, and added just a small bit more of that noise you get with imperfect inversions.
So I converted the lossless instrumentals to an MP3 with Foobar, and it was better, but still had that damned drum crackling! So I kept trying. I used OPUS, OGG, different bitrates of MP3, even AAC.
I have found that OPUS is the best at removing the drum overlap, I cannot hear any in fact, with OPUS.
So, my final guide is,
If you are getting cracking/crackling/overlap on drum hits in your inversions, then:
- Convert the instrumental to OPUS (128 Kbps) with Foobar2000.
- Use a software like Audacity to amplify it to a peak amplitude zero DB (since apparently OPUS auto declips to floating levels?)
- Export it as a WAV at the original sample rate (since OPUS only supports 48 kHz, I actually tried resampling the instrumental, and original to 48 kHz before converting to OPUS, but found that results in a WORST output.)
- Do your inversion (hopefully in Utagoe)
- Use whichever vocal model you like the output for best for cleanup.” - sausum
PS. Once, I ran into a similar issue. And I fixed it, actually similarly. I think I was trying to invert with an mp3 VBR, and the other file was lossless, so I converted it to the same codec and bitrate/V preset.
Yes, it wasn't perfect, but better.
I wonder if simply applying a cutoff at 20kHz wouldn't be a better solution. That's what Opus more or less does, plus upsampling to 48kHz.
- Despite the fact that separation is in 32 bit float, align inputs option in UVR uses something lower internally, hence the clipping may occur.
- As a better alternative to Utagoe and UVR’s Align feature, you can use paid Auto Align Post 2 (maybe even cheaper MAutoAlign).
"This is a substantially altered version of Utagoe, while it retains the underlying algorithms including in v3, I've added a looot of stuff, like themes, new waveform algos, centre+side extraction algos, new alignment algos, a console, codec support, and much more fun stuff.
Y'all go crazy, have fun." - Jake/commontea
“Multiple new superior algorithms for both WaveForm and Frequency domain, with added metrics for comparison. Output normalization, any file input, different sampling rates, the whole shebang. Although unlike UVR you still cannot export the changed target, you can only get the inverted mix still” - Vinctekan