haven't read the paper yet but God the audio is divine

> Experiment evaluations on popular LJSpeech dataset show that our proposed NaturalSpeech achieves -0.01 CMOS (comparative mean opinion score) to human recordings at the sentence level, with Wilcoxon signed rank test at p-level p >> 0.05, which demonstrates no statistically significant difference from human recordings for the first time on this dataset.

yeah really impressive

i've been trying fastpitch via coqui but it seems it'll take a long time for results in part because the default config has 2000 steps of learning rate warmup

glimmer has been trying flowtron on h but again long time

I think you should really implement dataloader speed optimizations

E.g get everything on ram first doing pickling

i barely managed to get coqui up and running i don't think i'm capable of optimizing dataloaders

tacotron2 implements a textmelloader class which is used during training so would it be possible to just pickle that

ye

I'll work on it later today

this is the issue where mels are loaded from disk in every step?

ye

How do i upload files and check if it works? I am not sure how much I send in or if a MP4 file would work.

It is clips for a test run of Gandalf from The Lord of The Rings. It is a few files of his voice, and mostly just his voice (some sort of backround music I think is in one) stringed together

is this supposed to be reference audio or a dataset

if reference audio then it might work not sure if dataset then you need to have it be specifically .wav files in 16-bit 22050hz mono, a few seconds per file, and also provide a transcript

its way too noisy to work as either

yeah needs to go through music removal

i recommend https://mvsep.com/ with these settings

or download https://github.com/anjok07/ultimatevocalremovergui

Just a bit more training and I think this will be an okay model to start. > Hey! I'm Kate Connors from Mirror's Edge, and you're on Uberduck!

I'm sure there's more lines in the newer version of the game but gotta start somewhere

Trained further and submitted

Hey, how much training data should be enough for a clear-sounding voice if trained enough? Atm I'm sitting on 12 minutes and 51 seconds of raw voice acting ripped straight from Disco Elysium waiting to be transcribed, should I add more? there's a total of 1 hour, 45 minutes and 37 seconds if I were to add all files I grabbed

Also, newbie question, should I add commas, apostrophes and exclamation marks to the transcripts?

(referenced from #768215837248716819)

seems good

Yea I'm sitting at 145 wavs @ 11 minutes 55 right now lol

Not sure how far I can go before I'll be met with troubles on a free collab account