yeah ffmpeg is very good

allows reencoding to 22khz s16le mono

I just dislike how some audio clips in this dataset are too high volume and some are too low so that’s my first priority

ehh for me I just don't care

What should I use to make a dataset on a Windows machine?

I usually use audacity and notepad.

i usually clip the files first in audacity (unless they're already clipped in which case convert with ffmpeg), make a list of them with command prompt, then play them in mpc-hc (because vlc cuts off the end) and use the media keys to go forward while writing in notepad

well this time I hope it sounds alot good

epoch 20 looking pog

So i'm going to go with lufs normalization technique

I'm not sure as to what the significant differences are between peak normalization and loudness normalization on the dataset when training

peak normalization sets the volume so that the loudest noise is some amount of db loudness normalization sets it so that the average is some amount of db

Ah thanks! @User have you considered making your own colab?

i have considered it but i'm not brain enough for it

i do have a fork of the normal one but the only code i've written for it is the dataset laoder

-or at least i thought so when i saw the heap of code that is step 2 but apparently that might have been copied from the repo?

yes

that is forked from the repo

oh step 2 of cookie's colab?

yes

from what i can see it just adds the extra info and the alignment graph which is something the regular tt2 repo does not have

and adds learning rate decay after 20k iterations

He also adds in torchmoji no?

dont think so

i dont see torchmoji being mentioned anywhere on the notebook

It might be a preprocessing step

Regardless I think @User you should try it out, i'm attempting to do everything from scratch (even though it's taking me a long time to understand everything)

i mean who knows, maybe

though if i do end up doing it it'll probably be talknet because talknet is based

that or some fancy version of tacotron