I wanted to try supervised training for future tacotron models I train, but do i have to label every file with it's arpabet sounds or just the words

you can put arpabet if you want to it

in my case for daffy duck i just used purely english words and let tacotron2 do the heavy lifting with arpabet

I meant rather, have every file in the dataset I have assigned to its respective arpabet phonetic so tacotron doesn't has to do it by its own

oh yeah that is legal

(i believe)

I mean it is legal, it just is very tedious lol

I just don't know how to label the dict for it specifically relating to data set sounds, i.e. the wav file "glue the sheet" has the arpabet labeling "G L UW . DH AH . SH IY T ." and etc

each word is separated by a period

so wrap them in curly brackets {}

oh each word seperately?

also make sure to check

show lexical stress

if youre using the cmudict website

yes

oh interesting, thanks for the tip, and will do

but when I want it to refer to the wav file, how would I have to specify it in the dict.txt file like the one you send

oh the dict is there to convert english words to arpabet

oh i see. I think I rather refer to something akin of a validation.txt of sorts, but I wonder if tacotron does use that too

i actually havent tested text with arpabet sequences beforehand and

oops

oh haha, damn yeah, sometimes it does mess up

though your daffy duck model caught the essence of the accent pretty close though so you're def on the right way

its because the dataset itself doesnt use arpabet sequences

im trying to look for a fix Right Now

yo good luck on that

wish i could help somehow, but im still new with that as well

trained my own model on a character I thought up for fun, and the results are good to mediocore, cuz I recorded the havard sentences instead of CMU artict lol

found a fix

also merged.dict.txt is included in this version

i realized i forgot to preserve punctuation on that screenshot

Now we need improvements over pronunciation. This should help.