Probably just gonna go up to 15 minutes for good measure

Sending help in a private message.

@TheTiagoFire quacked: How long of a model, either in regards to total length or to number of waves, do you think I should have before I can expect good results coming from the trained end result similar to this one that is speaking?

Getting quite late and I'm unsure if I should call it quits

dont use the bot outside of #842056330835656745

Ah my bad

Kind of went two in one there to show what kind of quality I was hoping to achieve with my dataset

That's a good idea

https://www.youtube.com/c/LGRBlerbs/videos - I think I'm gonna do this guy next. Fortunately his Blerbs channel has no music in the background.

personally i'd rather run a music remover on the main videos because they're scripted

Yep, thanks Hecko, it's not cleat and Speechtexter is having a hard dime.

*hard time

!!! making progress on training the japanese model to speak english

i'm surprised at how much of a japanese accent it has tbh. i was expecting it to unlearn it however, some of the other speakers have japanese accents (there are 8 speakers; 2, 3, and 4 have japanese-accented english, and 7 is the japanese dataset) so i suppose it makes sense that the accent is retained funnily enough, some of the american-accented speakers (the ones with very little data, mostly) have a bit of a japanese accent at times

it's having some trouble with naturalness, but it also only has 8 minutes of data, so. i think that's to be expected

this might be useful for damaged vocal audio since ik for a lot of sources, it's really hard to get the proper isolated vocals https://github.com/haoheliu/voicefixer

i'm testing it rn but the checkpoints are taking forever to download cuz zenodo

we've known about voicefixer for months

in my experience i've never found it useful though

oh sorry

i tested it with a few things; it doesn't sound great on mildly damaged audio it could help with something really bad, though

Tutorial says to run the training for 500 epochs max but should I amp it up to 1000? I'm fine with waiting two hours

He said 10000 would be 20 hours so I'm just doing the math there

For tacotron? Just stop it when it gets to a good val loss

About that

turned out to be useless since it doesn't even let me do it lmao

that's entirely too many epochs what the hell

yea it turns out the tutorial was outdated