VR ensemble settings
As for VR architecture, ensemble is the most universal and versatile solution for lots of tracks. It delivers, when results achieved with single models fail - e.g. when snare is too muffled or distorted along with some instruments, but sometimes a single model can still provide more clarity, so it’s not universal for every track.
In most cases, ensemble of only VR models is dedicated for the tracks when in the most prevailing moments of busy mix in the track, you don’t have major bleeding using single VR model(s) because it rarely removes that well vocal residues from instrumentals better than current MDX models, or with high aggressiveness it becomes too destructive.
Order of models is crucial (at least in the Colab)! Set the model with the best results as the first one. Usually, using more than 4 models has a negative impact on the quality. Be aware that you cannot use postprocess in HV Colab in this mode, otherwise you’ll encounter an error. Please note that now UVR 5 GUI allows an ensemble of UVR and MDX models in the app exclusively, so feel free to check it too. Here you will find settings for “only” UVR models ensemble only.
- HP2-4BAND-3090_4band_arch-500m_1.pth (9_HP2-UVR)
- **HP2-4BAND-3090_4band_arch-500m_2.pth (8_HP2-UVR)
- HighPrecison_4band_arch-124m_1.pth (probably deleted from GUI, and you’d need to copy this model from here to your GUI folder manually - if it will only work)
- HP_4BAND_3090_arch-124m.pth (1_HP-UVR)
(order in Colab is important, keep it that way!)
Or for less bleeding, but a bit more muffled snare, use this one instead:
HP-4BAND-V2_arch-124m.pth (model available only in Colab, recommended
*on slower Tesla K80 you can run out of time due to runtime disconnection, but you should get faster Tesla T4 by default on first Colab connection on the account in 24h.
Aggressiveness: 0.1 (pretty universal in most cases, 0.09 rarely fits).
Or for more vivid snare if bleeding won’t kick in too much: 0.01 (in cases when it’s more singing than rapping - for the latter it can result in more unpleasant bleeding (or just in some parts of the track). Suggested very low aggressiveness here doesn’t leak as much as it could using the same settings on a single model, but it leaks more in general vs single models’ suggested settings).
0.05 is not good enough for anything AFAIK.
high_end_process: mirroring2 (just ON in GUI)
(for less vivid snare check “bypass”, (not “mirroring” for ensemble - for some reason both make the sound more muffled), be aware that bypass on ensemble results with less vocal leftovers)
ensembling_parameter: 4band_44100.json
TTA: ON
Window size: 272
FlipVocalModels: ON
Ensemble algorithm: default on Colab (min_mag for instrumentals)
Other ensemble settings
- For clap leftovers in vocal stem, check out this ensemble settings.
- For creaking sounds, process your separation output more than once till you get there with this setting
- Also reported clean instrumentals with this setting
Make sure you checked separated file after the process and file length agrees with original file. Occasionally, the result file can be cut in the middle, and you’ll need to start isolation again. Also, you can accidentally start isolation before uploading of source file is finished. In that case, it will be cut as well.
It takes 45 minutes using Tesla T4 (~RTX 3050 in CUDA benchmarks) for these 4 models settings. Change your songs for processing after finishing the task FAST, otherwise you’ll be disconnected from runtime when the notebook is idle for some time (it can even freeze in the middle).
In reality, Tesla T4 maybe has much more memory, but what takes 30 minutes on a real RTX 3050, here might take even more than 2 hours and sometimes slower or sometimes slightly faster (usually slower). So you're warned.
**Be aware that these 4 model ensemble setting with both 500m models in most cases won’t suffice for the slowest (and no longer available in 2023) Tesla K80 due to its time and performance limit to finish such a long operation which exceeds 2 hours (it takes around 02:25h). Certain tasks too much above 2 hours ends up with runtime disconnection, so you're warned.
Also be aware that the working icon of Colab on the opened tab sometimes doesn’t refresh when operation is done.
Furthermore, it can happen that the Colab will hang near 01:45-02:17h time of executing the operation. To proceed, you can click F5 and press cancel on prompt to whether to refresh. Now the site will be functional again, but the process will stop without any notice. It is most likely the same case when you suddenly stop connection to the internet, and the process will still run virtually till you reconnect to the session. But here, you just don’t have to click the reconnect button on the right top. Most likely you have very limited time to reestablish the connection till the process will stop permanently if you don't connect on connection lost (or eventually if progress tracker/Colab will stop responding). So in the worst case, you need to observe if the process is still working between 01:45-02:17h of processing. If you see that your GPU has 0.84GB instead of ~2GB, you’re too late and your process is permanently interrupted, and the result is gone. It’s harder to track how long it processes when you already used the workaround once, and the timer stopped, so you don't know how long it is separating already.
Limit for faster Tesla T4 is between 1:45 and 2:00h/+ (sometimes 2:25, but can disconnect sooner, so try not to exceed two hours) of constant batch operation, which suffice for 2 tracks being isolated using ensemble settings above with both 500m models (rarely 3 tracks).
HP2-4BAND-3090_4band_arch-500m_1 (9_HP2-UVR) - I think it tends to give the most consistent results for various songs (at least for songs when vocal residues are not too prevalent here)
HP-4BAND-V2_arch-124m (2_HP-UVR) - much faster and can give crisp results, but with too many vocal residues for some songs (like VR arch generally tends to)
HP_4BAND_3090_arch-124m (1_HP-UVR) - something between the two above, and can give the best results for some song too (out of other VR models)
HP2-MAIN-MSB2-3BAND-3090_arch-500m (7_HP2-UVR.pth) - tends to have the least vocal residues out of the VR models listed above, but in cost of instrumentals not sounding so "full"
HighPrecison_4band_arch-124m_1 (I think not available in UVR, you'd need to install it manually) - can be a good companion if you only have VR models for ensemble
HP2-4BAND-3090_4band_arch-500m_2 (9_HP2-UVR) - the same situation, I think it rarely gives any better results than 500m_1 (if in even any case) but it's good for purely VR ensemble
_______VR algorithms of ensemble _______
by サナ(Hv#3868)
“np_min takes the highest value out, np_max does vice versa
it's also similar to min_mag and max_mag
So the min_mag is better for instrumental as you could remove artefacts.
comb_norm simply mixes and normalizes the tracks. I use this for acapella as you won't lose any data this way”
Batch conversion on UVR Colab
There’s a “ConvertAll” batch option available in Colab. You can search for “idle check” in this document to prevent disconnections on long Colab sessions, but at least if you get the slowest K80 GPU, the limit is currently 2 hours of constant work, and it simply terminates the session with GPU limit error. The limit is enough for 5 tracks - 22 minutes with ~+/-3m17s overhead (HP_4BAND_3090_arch-124m/TTA/272ws/noppr/~2it/s) so better execute bigger operations in smaller instances using various accounts and/or after 3-5 attempts you can also finally hit on better GPU than K80.
To get faster GPU simply go to Runtime>Manage session>Close and connect and execute Colab till you get faster Tesla T4 (up to 5 times). But be aware, that 5 reconnections will reach the limit on your account, and you will need to change it. It’s easier to get T4 and not reach the limit reconnecting, around 12:00 CET in working days. 14:30 o’clock it was impossible to get T4, but probably it depended on a situation when I already used T4 this day since I received it immediately on another account.
For single files isolation instead of batch convert I think it took me 6-7 hours till the GPU limit was reached, and I processed 19 tracks using 272 ws in that session.
JFI: Even 5800X is slower than the slowest Colab GPU.
Shared UVR installation folder among various Google accounts
Since we no longer can use old Gdrive mounting method allowing mounting the same drive across various Colab sessions - to not clutter all of your accounts by UVR installation, simply share a folder with editing privileges and create a shortcut from it to your new account. Sadly the trick will work for one session at a time.
Firstly - sometimes you can have problems with opening the shared folder on proper account despite changing it after opening the link (it may leave you on old account anyway). In that case, you need to manually insert id of your account where you want to open your link to. E.g. https://drive.google.com/drive/u/9/folders/xxxxxxxx (where 9 is an example of your account ID which shows right after you switch your account on main Google Drive page).
After you opened the shared UVR link on your desired account, you need to add the shortcut to your disk (arrow near folder’s name) and when it’s done, create “track” and “separated” folder on your own - so delete/rename shared “tracks” and “separated” folder and create it manually, otherwise you will get error during separation. If you still get an error anyway, refresh file browser in the left of Colab and/or retry running separation three times till error disappears (from now on it shows error occasionally, and you need to retry from time to time and/or click refresh button in file manager view in the left or even navigate manually to tracks folder in order to refresh), Colab gets changes like moving files and folders on your disk with certain delay. And be aware that most likely such way of installing UVR will prevent you from any further updates from such account with shared UVR files, and on the account you shared the UVR files from, you need to repeat folder operations if you will use it back again on Colab.
Comparing 500m_1 and arch_124m above, in some cases you can notice that the snare is louder in the first, but you can easily make it up using mirroring instead of mirroring2. Downside of normal mirroring might be more pronounced vocal residues due to higher output frequency.
Also, in 500m_1 more instruments are damaged or muffled, though more aggressiveness in the default setting of 500m_1 sometimes makes an impression that more vocal residues are cancelled.
(evaluation tests window size 272 vs 320 -
it’s much slower, doesn’t give noticeable difference on all sound systems, 272 got slightly worse score, but based on my personal experience I insist on using 272 anyway)
(evaluation tests aggressiveness 0.3 vs 0.275 -
doesn’t apply for all models - e.g. MGM - 0.09)
(evaluation tests TTA ON vs OFF -
in some cases, people disable it)
5a) (haven’t tested thoroughly these aggressiveness parameters yet)
HP2-4BAND-3090_4band_arch-500m_1.pth
w 272 ag 0.01, TTA, Mirroring
5c)
HP2-4BAND-3090_4band_arch-500m_1.pth
w 272, ag 0.0, TTA, Mirroring 2
Low or 0.0 aggressiveness leaves more noise, sometimes it makes instrumental cleaner, if you don’t care for more vocal bleeding (it depends also on your sound system how you are able to catch them. E.g. whether you listen on headphones or speakers).
But be aware that:
“A 272 window size in v5 isn't recommended [in all cases]. Because of the differing bands. In some cases it can make conversions slightly worse. 272 is better for single band models (v4 models) and even then the difference is tiny” Anjok (developer)
(so on some tracks it might be better to use 320 and not below 352, but personally I haven’t found such case yet)
DeepExtraction is very destructive, and I wouldn’t recommend it with current good models.
Karokee V2 model for UVR v5 (MDX arch)
(leaves backing vocals, 4band, not in Colab yet, but available on MVSep)
Model:
https://mega.nz/file/yJIBXKxR#10vw6lRJmHRe3CMnab2-w6gAk-Htk1kEhIp_qQGCG3Y
Be sure to update your scripts (if you use older command line version instead of GUI):
https://github.com/Anjok07/ultimatevocalremovergui/tree/v5-beta-cml
Run:
python inference.py -g 0 -m modelparams\4band_v2_sn.json -P models\karokee_4band_v2_sn.pth -i <input>
5d) Web version for UVR/MDX/Demucs (alternative, no window size parameter for better quality):
How to use this free online stem splitter with a variety of quality algorithms -
1. Put your audio file in.
2. Choose an algorithm. Usually, you really only need to choose one of two algorithms:
- The best algorithm for getting clean vocals/instrumental is selecting Ultimate Vocal Remover. Once you selected Ultimate Vocal Remover, select HP-4BAND-V2 as the "Model type".
- The best algorithm for getting clean separate instrument tracks, like bass, drums and other, is Demucs 3 Model B.
3. Hit Separate, and mvsep will load it for you. This means you can do everything yourself, no need to ask for other people's isolations if you can't find them.
6) VR 3 band model (gives better results on some songs like K Pop)
(I think default aggresiveness was 0.3)
7) deprecated - in many cases lot of bleeding (not every time) but in some cases it hurts some instruments less than all above models (e.g. quiet claps).
MGM-v5-4Band-44100-BETA2/
(MGM-v5-4Band-44100-_arch-default-BETA2)
/BETA1
Agg 0.9, TTA, WS: 272
Sometimes I use Lossless-Cut to merge beta1 and beta2 certain fragments.
Models from point 4 surpasses ensemble of both BETA1 and BETA2 models.
(!) Interesting results (back in 2021)
“Whoever wants to know the HP1, HP2 plus v4 STACKED model method, I have a [...] group explaining it"
Long story short - you need to ensemble HP1 and HP2 models, then on top of it, apply stacked model from v4.
Be aware that ensemble with postprocessing in Colab doesn't work.
Instruction:
1 Open this link
https://colab.research.google.com/drive/189nHyAUfHIfTAXbm15Aj1Onlog2qcCp0?usp=sharing
2. Proceed all the steps
3. After mounting GDrive upload your, at best, lossless song to GDrive\MDX\tracks
4. Uncheck download as MP3, begin isolation step
5. Download the track from "separated" folder on your GDrive. You can use GDrive preview on the left.
1*. Alternatively, if you have a paid account here, upload your song to: https://x-minus.pro/ai?hp
Make sure you have "mdx" selected for the AI Model option. Wait for it to finish processing.
2*. Set the download format to "wav" then click "DL Music." Store the resulting file in the ROOT of your UVR installation.
6. Use a combination of UVR models to remove the vocals. Experiment to see what works with what. Here's a good starting point:
HP2-4BAND-3090_4band_arch-500m_1.pth
HP2-4BAND-3090_4band_arch-500m_2.pth
HP_4BAND_3090_arch-124m.pth
HP-4BAND-V2_arch-124m.pth
7. Store the resulting file in the ROOT of your UVR installation alongside your MDX result.
8. Finally, ensemble the two outputs together. cd into the root of your UVR installation and invoke spec_utils.py like so:
$ python lib/spec_utils.py -a crossover <input1> <input2>
the output will be stored in the ensembled folder
9* (optional). Ensemble the output from spec_utils with the output from UVR 4 stacked models using the same algorithm
Ensemble
spec_utils.py allowing ensemble is standalone, and doesn't require UVR installed in order to work. It accepts any of the audio files
mul - multiplies two spectrograms
crossover - mixes the high frequencies of one spectrogram with the low frequencies of another spectrogram
Default usage from aufr33:
python lib/spec_utils.py -o inst_co -a crossover UVR_inst.wav MDX_inst.wav
https://github.com/Anjok07/ultimatevocalremovergui/blob/v5-beta-cml/lib/spec_utils.py
Custom UVR Piano Model:
https://drive.google.com/file/d/1_GEEhvZj1qyIod1d1MX2lM6u65CTpbml/view?usp=s
______________