"Mmh" noises not being properly converted/translated

[umbral crag]

Hi I'm trying to improve my model. Its unable to make some basic sounds, everything else works perfectly fine and it sounds great otherwise.
The dataset is very clean but does not contain any prolonged mm sounds (since some consider it to be "unclean"), should I try adding some or how would you go about this?

[maiden vigil]

yeah, you should definitely try to add those then

in RVC, you should always add what you want to be replicated, and it will replicate that. when you don't give it certain data, it can't do that sometimes

[umbral crag]

I'm still a bit unsure about how it estimates what sound I'm trying to make.
I guess if my entire dataset was mmh and umm, it would be perfect at it but how it decides if both exist, humm.
In any case I will try to fine some samples. Thanks

[maiden vigil]

You're welcome! Let me know how it turns out

[umbral crag]

Hi so I added a ton of different samples some of which containing M and U sounds (it stuggled with both before).
Unfortunately even though the additions now make up about 10% of my dataset it didn't improve the models ability to translate those sounds at all.
I'm assuming that the issue is not with the model itself but the voice interpretation, that it doesn't properly recognize me making those sounds :/
I EQ'd my mic a bit to lower the bass, which improved the overall translation quality but not that of M and U sounds.

Could you send me an audio file where lonely M and U sounds are succesfully translated? Doesn't matter what model or with your or my voice. I'd like to hear if and how well it works for others.

Here is one of the samples of which I hoped it could improve the model (original voice, not translated). Just so you know what I tried