not wrong just a different one

if it sounds bad it's probably because the preview uses a garbage vocoder

try the synthesis notebook linked somewhere around the training notebook you chose

what differences of pipeline and non pipeline?

main difference is that pipeline uses arpabet by default which means its models will pronounce things better

but also it's fiddlier to use

think by default you have to move the model to drive manually after you're done training

what symbol does pipeline uses?

nvidia_taco2 by default i think?

i dont think this is good

I am trying to train it to Hifi-GAN

i dont know if something wrong with my path

wavs.zip are basically wavs

i don't think it's possible to train hifi-gan for a pipeline model currently

though i think i might be able to hack something together once i get home (aka in many hours)

many thanks

230 wavs, 200 epoch

what do you think would make it better next time? more wav input or more epochs?

lol

427 wavs 100 epochs because 11 batch size 3800 Steps

i trained this using Pipeline

if you can get more wavs then absolutely do but tbh this is already very good, if you wanna get rid of the roboticity then use the hifi-gan finetuner notebook

yes, theres actually around 500+ clips in total. (extracted from the game) but wanted to try half first. If i put them all in would it improve it? and should i keep it at 200 epochs ? but thanks! and perhaps ill try out the finetuner as well

absolutely put in as many clips as you can, except the ones where the character shouts or grunts or such

and i recommend looking at the validation oss, with this size of dataset around 0.14 should be enough according to dawn

yeye!

ooo ok

but even better is to make a backup every so often and then compare them

i've had a model trained for 40 epochs sound subjectively better than after 120

oooh