Any advise on how I'm training my Ozzy Osborne model? These are the results of him singing his training data (TalkNet)

Why does starlight3 produce that robot synth while breathing

Noisy data really makes it that worse huh

nono 0 is the noisy data one

Yeah I was referring to the first audio file in that second sentence

0 sounds way worse with noisy data didn't know it was that significant

i mean it was uh

40% clean, 20% noisy, 40% very noisy

Why is 3 the "good" one

because it was trained with properly processed data

but idk about the actual quality

my ears refuse to work asdgfbn

Did you use hifi gan?

yeah, not finetuned though

could you send a sample of your data

so that's talknet right?

nope

those are all tacotron

because the person who requested them wanted tacotron

oh the duration prediction artifacts sound much worse than I got used to from tacotron2

?

?

what are duration prediction artifacts

I’m guessing it’s how models predict the next frame in the output

According to Cookie “tacotron2 predicts the next frame using the previous frame. When training, tacotron2 is always given the previous frame from the original audio, so it only have to predict 0.0125ms into the future”

Not sure if it’s a duration prediction artifact tho

talknet has a duration predictor

tacotron doesn't have one yeah

either way what could an artifact therefrom be

@User I used a couple of synthesis notebooks in colaboratory, and make 4 models, none works. What should I do?