2024 "BS-Roformer" spectral algorithm - Has anyone used it?

[Tex]

There's a new algorithm out there on the spectral sites has anyone been using it?

It processes with a new method that supposedly gives better results than the others.

[MP3J]

give it a try here... https://mvsep.com/en, select from the drop down menu

[Lord Reith]

The title doesn't exactly inspire great confidence... "B.S."

I tried it but as usual it's a case of swings and roundabouts. A lot of these new algorithms work really well on one type of music and not well on another. Demucs4 is rated the top algorithm but it is very hit and miss when it comes to vocals and bass. Often it'll capture a whole lot of other stuff with the voices and miss entire passages of bass guitar. The best vocal algorithm is for sure the MDX one near the top of the list. That nearly always gives a perfect result. UnmixXL can get bass guitar where every other one fails and Demucs3 or 4 are best for drums depending on the source. So you really need to be using a lot of different ones depending on the song.

[Tex]

Yeah, can't really see the algorithms getting too much better than they already are.

If you are largely overlapping them in the mix it doesn't really matter if you have 100% separation anyways.

[Tex]

BS-RoFormer relies on a band-split module to project the input complex spectrogram into subband-level representations, and then arranges a stack of hierarchical Transformers to model the inner-band as well as inter-band sequences for multi-band mask estimation.

Umkay.

[Bambam34]

It Was really Good For Extracting The Vocal To Julia.

[Lord Reith]

They do seem to have reached a ceiling for the moment. The big jump was from Spleeter to Demucs3 but since then it's been pretty incremental. And none of them produce stems which are clean enough to be taken for original tape tracks. MAL seems to be able to do this although since that is a black box there's no way to know whether it is an automatic process or whether they are applying extensive post production manually. The main advantage the MAL team have is unlimited funds I guess.

However having said that, there is no real need for stems to be perfect if they are being recombined back into a mix. In that case the artefacts cancel out.

[EdinSc]

what does this cost? has anyone tried it with live tracks?

[Lord Reith]

It's free.

[Tex]

I tried a vocal extract of "Old Brown Shoe" and it seemed slightly better than the last time but that's the worst vocal to attempt to pull out.

The difference I heard was mainly a little less lead guitar artifacts on the vocal track.

The new algorithm seems to be only be a lead vocal separator which means if you want more tracks you'd still have to run it through the other algorithms. I also noticed it severely wavers out of time with older extracts.