Best workflow for RVC dataset: dereverb → lead separation or lead separation → dereverb?

[wintry sage]

Hey everyone, I have a question about the recommended workflow for preparing vocals for an RVC dataset.

Which order generally gives better results?

A) DeEcho/DeReverb → Lead Vocal Separation
B) Lead Vocal Separation → DeEcho/DeReverb

I want to follow the workflow that experienced RVC model creators usually use.

Additional context:

When I dereverb first, I get a cleaner and better overall tone, but the output keeps some adlib reflections, reverb traces, and echo remnants.

When I separate first, most of those reflections disappear, but the vocal tone becomes more unnatural, less isolated, and somewhat degraded.

If anyone who has trained multiple models can share what’s worked best for them, I’d really appreciate it!

[celest leaf]

i tend to go with A, gives best results

[wintry sage]

And do you encounter any reverb traces or echo?

[celest leaf]

no, havent

[night kite]

I would also try this method:
Separate vocals from instrumental -> dereverb -> deecho -> denoise -> separate lead from backing vocals -> denoise
Previously, I tried this method (B):
Separating the vocals from the instrumental -> separating the lead from the backing -> dereverbing -> deechoing -> denoising
Compared to (A):
Separating the vocals from the instrumental -> dereverbing -> deechoing -> separating the lead from the backing -> denoising
The results from this method sound way better than A and B.

[wintry sage]

Thanks, will try this. What deecho/dereverb models do you use? I use the UVR DeEcho-DeReverb model and it mistakenly hard-cuts certain parts by confusing them for reverb tails creating like a bounce effect.

Also what denoise model do you use?

[night kite]

First of all, I've been using these models on MVSep.
The models I've used for dereverb are version 2 of Melform Roformer dereverb model by Anvuew and the BS Roformer room dereverb model also by Anvuew. I also used UVR de-echo-normal with an aggressive level of 0.3 for deechoing.
Lastly, I used aufr33's DeNoise model, the aggressive one particularly. You might find it as "Mel-Roformer-Denoise-Aufr33-Aggr".

[wintry sage]

I also use MVSEP.

[night kite]

Ah, I see.

[wintry sage]

I saw it but didn't try because it's named as reverb removal not dereverb.

[night kite]

Reverb removal is another name for dereverb models. By dereverbing, you're removing the reverb from the audio.

[wintry sage]

I also tried the deecho normal but didn't like it because it removed a lot of brightness and clarity.

I use both models on 0.1 aggressiveness

To avoid any potential damage

I didn't use the room dereverb by Anvew since it's only trained to remove room reverb type

[night kite]

I see. From what I heard, UVR De-echo compresses the spectrogram of the audio's sample rate to 32k. Your only alternative to deecho would be using Clarity VX by Waves.

Personally, I'd use the regular dereverb model and then room dereverb before deechoing.

[wintry sage]

Thanks for letting me know that. I'm going to try the Melform Roformer dereverb by Anvuew.

[wintry sage]

It doesn't deecho. It only does dereverb.

For deecho you can use RX or the UVR deecho model

These are the only two options.

As far as I know

[night kite]

No, I'm referring to de-echo-normal, not deecho-dereverb. The latter is what you're referring to when dereverbing.

[wintry sage]

Oh ok, now I understand

[night kite]

Yes, de-echo-normal does de-echoing only. Deecho-dereverb can do dereverbing only.

[wintry sage]

I heard that the Melform Roformer dereverb by Anvuew drys out the vocals too much so I'm thinking of taking the reverb output and blend it at very low volume to compensate

[frank fern]

the aggressive model is better imo

[frank fern]

it's quite good but doesn't remove echo

[wintry sage]

Does the Melform Roformer dereverb by Anvuew do deechoing?

Oh ok

[night kite]

What aggressive level do you recommend when using the aggressive model of UVR de-echo?

[frank fern]

try the Sucial model for some heavy echo

[wintry sage]

What's your go to?

Is using different settings/models for different tracks creates inconsistency in the model?

What's your usual workflow for removing reverb, echo, backing vocal, and noise?

[frank fern]

neither of the models we're talking about are perfect for all cases, in the end you'd sometimes have to try several models (and workflows) and combine the best parts of them in post-editing

i.e. pick the parts having no leftovers and being full

ik that's the effort that'd come to

[night kite]

My workflow is as follows:
MelBand Roformer (becruily vocals high fullness) -> MelBand Roformer (Reverb removal by anvuew (version 2)) -> BS Roformer (DeReverb room by anvuew) -> UVR De-echo-normal (aggressive level: 0.3)) -> Denoise model by aufr33 (aggressive) -> MelBand Roformer (MVSep Karaoke by gabox) -> Denoise model by aufr33 (aggressive)
It probably may not work as expected.

[wintry sage]

But that will require a lot of work. I want a workflow/model that works well with all to apply to all tracks

[proven wave]

Anyone has their own workflow, but each have trade off and might not be 100% perfect for everyone. While my own purposed workflow might sound simple, that one is supposed for use in AI cover, remastering and not for train an actual voice model.

It's more of trial-and-error rather than the quick or exact solution. Rather, you can try all methods above, and compare them by tools and ears.