How i rescale the audio?

[urban wing]

I want to convert my 32k dataset to 48k. i saw a github but i dont remember the name, someone know?

[mighty lily]

I think @woven sonnet or @kind nimbus might know

[mighty lily]

Para que lo necesitas exactamente?

[woven sonnet]

you mean you want to upsample your audio?

if you mean that i wrote a small guide on it once lol

https://rentry.org/upscaleaudio

you dont really need it for rvc tho

if thats what you are looking for, this was probably the github you saw

https://github.com/brentspell/hifi-gan-bwe

[restive otter]

Why should one do that? The audio dataset is gonna be converted to 16khz anyway

[woven sonnet]

true, thats why im saying this doesnt really make a difference for rvc

its useful outside of that tho

[solar island]

Are you referring to resampling the sample rate or upscaling it by recovering lost high frequencies?

[solar island]

that's the part I don't understand about rvc, if it limits to just 16khz then why do we have 40k and 48k pretrains?

[bronze basalt]

for F0 iirc, but a proper 48k dataset will still output full 48k during inference

older model, ignore the HF meme around 22 kHz

source audio

i artificially cut it down with hpf and lpf though and did as a 32k anyways

[timid tangle]

It does that for hubert to extract features of speech, sort of understand what the audio is saying, marks every word and sound basically and that's it, sort of like an STT but a lil bit different, gan uses 32/40/48k cuts to actually train the voice model, so no, it doesn't downscale audio to 16k to then use it to create a model, it uses it to just mark words and sounds with the hubert model. Basically RVC has different modules, first is the slicer, when RVC cuts audio into small pieces. It creates two dataset folders out of your dataset, the f0 one for pitch and sound generation which will be sampled to whatever config u choose, and the dataset for the hubert_base which works only with 16k sample rate audio and extracts features. Then the gan module will just read allat information regarding pitch and features that the f0 model and hubert_base model created and it will create the Generator model and Discriminator model which will then try to fool each other to make the sound model sound like the dataset you gave it. What's funny you can skip the hubert part or ruin it so that it fails and then the model will sound like it's having a stroke mispronouncing every word, shit sounds funny af.

[restive otter]

https://tenor.com/view/fact-do-you-single-to-back-that-up-gif-21443611

[timid tangle]

If it used downscaled to 16k audio to build models they all would sound 16k and low quality

[restive otter]

I've looked at extract_feature_print.py, extract_f0_print.py, no code referencing 0_gt_wavs dir