How to increase vocal "range" on merged voices?

[sick bay]

Hi all, I'm creating a custom merge of some voices to get the unique voice I'm looking for, but one issue all my iterations have is the inability to hit a high note. For example if I do a high-pitched "eeeeeeeeeeeee" it crackles and cuts out. Same with a laugh-- it comes out as crackly and more like the mic is messing up.

Do I need to find a model with a better range? Is there a way to merge multiple models and include a model that has the better range without destroying the existing sound? I'm sure it's complicated but I'm new to the scene, so looking for some guidance. Thanks!

Using a local install of RVC to train or create, and MMCV to modulate live voice and for merging models.

[alpine mist]

I have heard that merging models kinda screws up singing models but I'm not sure... Maybe try using the fork of RVC you're using to merge? Haven't really tested the feature so I don't really know much about it

[wicked aurora]

Merging models degrade quality and affect overall range

Even if you pick 2 high quality ones it will degrade a lot in quality

[sick bay]

Aw man. It's hard to find the exact voice I'm trying to create. But that's good to know-- thank you.

[bright barn]

someone tested before and it somehow improve the pronunciation of words and way more natural after he merge different models, not sure how he does it, it could be different dataset merge into one index file

[wicked aurora]

in some cases, if you fuse 2 models of the same language

it might help the pronunciation

but overall you will get artifacts

[bright barn]

well I have done it before also, but like it doesn't sound like an artifact, more like bleeding

it was an odd test because it was 2 models 1 with english and one with anime voice

some words kept switching back and fourth

[twilit slate]

haven't tried the model merge feature, if it has the same/similar effect as a single model trained from dataset joined from datasets of each model you want to fuse?

[winter osprey]

I've seen a video of an youtuber testing that model merge, iirc if you mix an female voice with an male voice, on higher pitches it will sound like a female and on lower pitches will sound like male.

[crimson forge]

the dataset needs to also cover that high range

otherwise the model completely just gives up

[winter osprey]

I always wondered how much mins of X voice you need to make a good og model.

I mean, if you put 2 different persons in a same dataset.

[crimson forge]

u need good gradients and variety

[winter osprey]

For example, if i wanna mix Gura's and Ninomae Ina'nis voices in a same dataset.

These two in their streams actually got a good vocal range.

[sick bay]

Hmmm, but merging too many models would ruin any clarity, correct?

[pseudo timber]

Ayo? @sick bay level 1 !!!

[sick bay]

I merged like 5 models together, so that's probably the issue. Seems I'll need to stick with one or two if I want to keep that range

[crimson forge]

well then u are simply including wayyy too much

the gan doesnt know how to generalize after a certain point

therefore u should only merge trained models

and not already merged models

[sick bay]

Gotcha. Dang it!

I'm happy with the voice itself but those high pitches and laughs kill it

[crimson forge]

yeah cuz the dataset didnt have any higher pitch or laughing it

[sick bay]

Well presumably at least one of them did but 5 smooshed together may have caused problems

[crimson forge]

it did

its way too much

u should never merge a merged model

[sick bay]

Well the models themselves weren't merged, I merged 5 unique models all at once

[crimson forge]

Yeah that just confuses the model after a certain point

Not saying it can't be used and such

But it's always a good Idea to merge as little as possible

[sick bay]

Good to know. I'll try 1-2 models and see what I get. Thank you!