what symbols?

polish

which includes arpabet for whatever reason

interesting

so i just ran g2p on the english data

ok gotcha

have you done much IPA training?

extremely dirty but i didn't feel like making a proper phonemizer

I did something similar training japanese to speak in english I just got a bunch of english speakers, plus a japanese speaker transcribed in arpabet

nice

I guess I just assumed IPA was a requirement for getting good results there

ipa would help a lot since it has phones that arpa does not

(and arpa is Engineered for english)

problem with ipa: polish /t/ is different from english /t/ because the latter can be [tʰ] sometimes

It certainly has a japanese accent but it worked pretty well I need to finish training it though A friend of mine started english->japanese too, by just transcribing the english with a japanese phonetic system

maybe a small language embedding would help

probably yeah

and then it could be switched around for accents

yeaaaah

i do think that's what google does

where've you seen that?

grazie

yep > (3) Incorporating an explicit language embedding to the input, which enables moderate control of speech accent, independent of speaker identity, when the training data contains multiple speakers per language.

They're saying words but idk wtf it means

also i like how they handled stress instead of having separate embeddings for AA0 and AA1 and AA2 they just have an embedding for AA and then add the embedding for 0 or 1 or 2

getting closer to daffy duck saying twoja stara zaiperdala

All I'm understanding is they throw shit in, tell it sound natural in 9ther Languages and shit happens

yeah

mostly