Make it sound realistic?

[maiden egret]

I am not able to make any model (female) sound convincing, is there a way to fix this? No I don't mean tweaking settings to find that sweetspot or looking for an outdated crepe-tiny version or whatever. I am looking to make my voice sound very convincing like a girl, but no matter how much I try it's almost impossible and people can tell.
Is the technology not there yet or can I fix this somehow? I thought about training my own model but apparently you can't make it "fit" to your voice so it's kind of pointless

I want to be able to go into a random discord or valorant game and people think I am a girl just based off my voice

Any help appreciated!

using: w-okada, rvc

[vital hollow]

rmvpe is the best voice convertion atm, just make sure you have a very clean dataset rid of background audio, echo, reverb, other voices than the character you want, and rid of any other sounds that is not from the chracter you want

[maiden egret]

I have done this, and I've tried gura models from 100 epochs up to 700 epochs and lots of them are very popular, this is not an issue of dataset

But your answer does not seem to deny the fact that you can create a convincing model that no one will notice is RVC? You don't seem to have this issue

[vital hollow]

yup, i personally use 1 hour to 2 hours of dataset for my models and most of them sounds realistic enough for people to not notice instantly

also the voice changer settings helps with the realism part of the voice

[topaz sage]

natural sounding speech? You need a dataset of more than 20+ minutes that is consistent through out the entire audio. The voice must be expressive have low and high pitch variations and so on.

and try to tweak the pitch settings that fits your voice

[maiden egret]

How do you train your models? Do you just use the aihub tutorial?

[vital hollow]

i dont really use the tutorials here but how i do mine is i get dataset, clean it, label it(save as multiple small clips of the voice), then train with Ov2 pretrain

[maiden egret]

pretrain?

[maiden egret]

dont leave me

[vital hollow]

ya pretty much an already trained model that your dataset will go ontop of to sound better faster, pretty sure you can train without it but it may suck

[honest lichen]

How can we do that? I have one more question: if I use a model with an American accent as a base and the dataset is my voice with a Japanese accent, will it be able to create a model with less japanese accent?

[vital hollow]

No, depends on how long you train but most of the time it should be somewhere in between

[honest lichen]

May you guide me?

[torpid yacht]

You trying to voice troll?

[light zephyr]

Also, the dataset must be HQ, very clean, extense, using a HQ pretrain, etc.

[torpid yacht]

Can get away with just using decent speech models if you merge

[maiden egret]

Catfish yeah

[maiden egret]

@vital hollow help him

[stable walrus]

Ayo? @maiden egret level 3 !!!

[maiden egret]

Where do you find a HQ pretrain even?