I have separated acoustic guitar tracks using Moises stem separator. It's not super clean. There are some digital artifacts and slight drop outs, usually at the end of notes.

Anyone know of a plug-in that would remove the artifacts and perhaps smooth over the dropouts? Thanks in advance.

 

Manual work through Spectralayers or RX is the only way. Or just recreate them.

 

Something like dxRevive for sounds other than vocals/dialogue would be neat.

 

All of the above - Spectral layers will possibly be the most accurate but it is also time consuming to get it right.

 

You can try to mask and smooth the artifacts by using a reverb send to hide some dropouts.

Before trying to fix anything manually, I’d recommend separating it again with a different method. Use UVR5 with htdemucs_ft, or an ensemble with htdemucs_ft and RoFormer‑FT. There’s a sample mode in UVR5 where you can test 30 second segments to check your settings. Both models can be downloaded directly in UVR5.

Starting with a cleaner separation will save a lot of time, because fixing artifacts after the fact will mostly be manual work.

 

Agreed 100% its crazy how good the quality can be with these models. What I do is run a few passes, using a few different models, then I load those stems into ableton and create a comp from them. As one pass may leave an artifact somewhere, but another pass hit that particular note just fine. Doing this ive been able to comp almost perfect takes.


Other than that, there’s not much else you can do besides the tools mentioned above. Even iZotope RX can work sometimes. Depending on the length of the artifact and the source material, you can also get away with zooming all the way in, cutting around the click, and deleting that chunk. From there, either move the tail of the note over and crossfade, or copy/paste from a clean section just before the artifact.

If you do that, zoom in extremely close and make your cuts at zero crossinggs so you’re not introducing clicks. Try to grab a replacement that matches the waveform shape and tone as closely as possible, and keep it relatively short. When you paste, use shortish crossfades on both sides to blend it in without smearing the vocal. This technique is tedious, but has saved my ass for decades now. Give it a shot.

 

most everyone learns the hard way about the importance of starting samples, and this is not much different.

Without knowing what kind of guitars (and their parts) were recorded, answering this sort of question is not a cookie cutter kind of thing. Every track separated, different models, and so on, can give really varied results. And that is all without any consideration what is the intended use of the separated stems. I've ended up using a number of the old tricks people used to use to make unauthorized remixes. You have EQ/filter separation which can work ok. Another good one is to find an instrumental version of the track and then doing the old inverted phase cancellation trick to keep only the vocals. In a situation like this with just two guitars, you might get away with using a splitter or multiband tool. The goal is to just strip out as much as you can before any pass through UVR5. Sometimes even just loading the sample into Serato Sample 2 can get rid of a bunch of stuff.

It's all pretty much about removing as much as possible, anything which might "confuse" the separation tool. But the long story short, is you really can't expect perfect from any of these tools yet. "Fixing" the results how the OP has asked, can often be more time consuming than it is worth for the intended usage.

 

I second this one. UVR has plenty of models and also allows to keep format and quality (I assume you're using the free tier of Moises)

If you don't want to run into installing UVR you can use MVSep which uses the same models or something like UVROnline