Generated index file ruins quality of inference

[raven latch]

Using my model without an index generates a better quality file compared to using the generated feature index file. Using the index, the voice sounds more like the pretrained model and ruins the output. The index adds a sort of lisp. I may be missing a key step here.

Here are some examples that show what I'm talking about

[copper prairie]

ur dataset lacks diverse data therefore when using index u are just limiting the output of the model

plus ur input audio isnt helping either

[raven latch]

How diverse should my dataset be? What range should I provide it in order to provide better output?

Or should I be removing things like similar takes in favor of unique data in my dataset

[steady harness]

Ayo? @raven latch level 1 !!!

[raven latch]

Because if so, then I don't mind having a much smaller dataset with more unique takes. I just have a lot of raw inputs that are similar in takes

[copper prairie]

First and foremost u need better quality audio. U also need like 30 mins of audio data that's diverse in order for the index to help

[raven latch]

I see, for the dataset I'm using raws rather than mixed vocals as I want to most accurately recreate the voice. Should I aim to use mixed vocals instead rather than raws?

[copper prairie]

Raw all the way

[raven latch]

Perfect, that's what I thought. Thank you for your help, any major pointers other than what you've said? I'm just using sessions with raws so

[copper prairie]

Use a batch size of 8

And use rx10 for de nosing and other audio cleaning

[raven latch]

Perfect, thank you so much!!

[copper prairie]

There's more u can do but I won't explain that now

I will later

[raven latch]

Sure, mind if I add you so I can keep in contact?