Manual ensemble
- Make sure you use lossless files of the same sample rate (or when lossy, potentially from the same format and source), otherwise the files will begin in different places, e.g. if one is lossy or in different lossy format, and then the process will require manual alignment in DAW beforehand, or even impossible in case of different SR till you resample.
- Don't try to use more than 5 models for one ensemble (SDR will decrease above that)
- Max FFT and Max Spec are the same (one is just inside the ensemble script/Colab [below], the latter is in UVR).
- MVSep uses normalization if clipping occurs. If you use such results with e.g. Colab separation, “It's like giving a bit more weight to the non normalized one over the normalized one”. So use files normalized for all stems in the bag, or all non-normalized.
32-bit WAV for premium users on MVSep gives non-normalized separations with reversible clipping, if you just turn down their volume manually in DAW (peaks in 32-bit float are not clipped when turning down the volume because of higher dynamics).
UVR has an optional normalization too.
It “reduces amplitude below the threshold across the whole track. It seems creating a lower base volume which is obviously not ideal” although, it might be better for the ensemble process if you use non-clipped results (at least if you don’t use 32-bit float outputs), but clipped might appear closer to the sound of the mixture.
It all happens, because vocals with instrumental when mixed together have frequencies which cause cancellation which results in decrease in volume in relation to just separated instrumental stem.
To avoid normalization on MVSep, you can also decrease the volume of the mixture before separation, but it affects separation results to a certain degree, but at least you know the value of normalization to restore it to the original volume, but working in DAW in 32-bit float, and manipulating volume and peaks freely.
How to ensemble ready separations?
(manual ensemble)
- On your already separated files in UVR5 under "Audio Tools" menu (or the regular ensemble for local models inference only), or also you can also use
- DAW method (weighting/mixing each stem manually in real-time), or alternatively you can also use
- Use Ensemble GUI (local) or Manual Ensemble Colab
Auto ensemble on sites
(performs separations, and ensembles automatically)
- MVSep presets editor can be used to create a node for the ensemble of your choice (“If 2+ algorithm nodes, it is premium only”), avg used by default. Some presets are already shared by community
- nextgen.uvronline.app allows ensembling of specific vocal and instrumental models with specific algorithms when you're logged in and enable show post processing menu in Advanced settings>Home screen for premium users