Trained Models Always Muffled Sounding

[stiff meadow]

I've been following the AIHub Docs guide for training a model: https://docs.aihub.gg/rvc/resources/training/
But my models always turn out muffled, and I'm not sure if it's my data or my settings
To try and isolate what's wrong I even tried to find as clean an audio source as possible for any voice to train, which led me to the Nijisanji Selen audio files on Internet Archive
I've removed silence as best as possible, there's no background noise as far as I can tell, there's no music, but it's still muffled

Here's an example of one of the tracks used in the dataset, I merge them all into 1 file that's 2 minutes 32 seconds long
I know that's not a ton of data and I have more I could merge together, but I feel like it shouldn't be this muffled even with that small of a dataset

I'm on codename-rvc-fork-4, using basically default settings except for Batch Size which was set to around 6, and me enabling envelope loss
I've already tried other optimizers, different pitch extraction algorithms, more training, less training, and it's always a fresh train each time with the pretrain from here #1339155300720054316

[stiff meadow]

I'm not even sure if "muffled" is the right descriptor, but it sounds like a compressor or something making it never sound clear when talking
It's so thick sounding that you can literally see how busy it is when doing an inference
The first image is my 40k trained model, the second image is me using a 32k Pomni model from this Discord

[brittle river]

Doesn’t sound muffled to me

[stiff meadow]

I figured it out, I think it was likely the audio quality in the end
When looking at even the clean samples from Nijisanji Selen, they had a buzz at around 18khz that made it into the final model, so they weren't as clean as expected even though they come completely isolated
I also swapped some settings around randomly, so at this point I'm so far from my starting point that I can't definitively know what made it better in the end

[brittle river]

Are the audio images less thick now