I'm like recording an own character I own though lmao, I had some genuinely decent results when I experimented so I decided to fully commit on it now

i've given someone a twilight sparkle dataset to train in 44k but i haven't heard back from them

Damn they ran away

how would one go about training in 44k?

i hope just change the funny number

i do have my street fighter datasets and a patty wagon dataset in 44k

fastspeech2 is also smaller and hard-aligned so probably worse too and please don't train for too long, else it'll obsess over the specific sentences you trained it on at the cost of everything else currently used notebooks are at the top of #841437191073955920, lower down are modern ones which i don't think you can submit results from

y'know what i'mma start training this

just to see if it works at all

why does linuck have to be so pissy about parent directories not existing like if you're gonna make the

twi

directory why not make

speakers

too

heck audacity makes directories by itself when i open the save dialog and it last was in a dir i've since deleted

interesting that pipeline has

sample_rate

in the config but also adds

sampling_rate

changed both just in case

well, talqu by haruqa has models that go up to 48khz so im sure its possible to just train models on 44khz

just takes more power i figure

from reading the ppp threads i recall that higher hertz is less stable

well that's, an interesting result; (loud warning near the end)

and when slowed down to 22k

gonna try out some things training on a small subset

also when will i learn to normalize my datasets

so here's a fun dilemma

setting sampling_rate to 44100 produces audio that is twice as fast as it should be whereas setting it to 22050 produces twice as slow audio

clearly the logical thing to do is 22050 * sqrt(2)

,,,it,,, sorta worked?

pitch is correct(?) but it's slower than it should be

@User what the hay is up with the sampling_rate option (i'm trying to train on 44k audio)

Yes

jesus

what a mess

Oh jeez that's A crunchy looking graph