I keep getting the damn K80... would it be more efficient to just run this on my own 3070 locally?

No

Does anyone know anything about using talknet for something other than english?

hey there, yeah you should look into local training

how much VRam do you have on your 3070?

8gb

Hey, does anyone have any resources on how to work with arpabet? I've heard it can really improve models but I'm not sure where to start. How do you go about converting your transcribed dialogue into arpabet? Is it all done manually or do you use the cmudict to translate the known words first, leaving the unknown words to be translated manually? If so does anyone have a notebook for doing this? I feel like it's a waste of time to manually translate the words that are already in the cmudict. I guess they already get translated by Tacotron when you have the submit a purely english transcript?

there's a notebook at the bottom of #841437191073955920 that does it automatically (but it's not the best and i don't feel like fixing it because pipeline training will replace it soon™ anyway)

once the model's up though you'll need to ask a site mod to enable the arpabet

When you say "it's not the best" what do you mean? That it doesn't work for all words or the conversion is poor?

it doesn't take context into account

sorry, could you give me an example

so for example "did you read this" vs "have you read this" it'll do the same pronunciation

ah ok

pipeline uses something called g2p which handles this pretty well

but the current notebook only has dictionary lookup

Does the pipeline notebook use Tacotron2?

yeah

but the model is formatted differently and currently not accepted on uberduck

ah, that makes sense

and i have no idea when it will be accepted

but you think the pipeline produces better results?

i haven't done a thorough comparison

all i know is that it does arpabet by default (which is good), and it has a different base model (which might be good?)

Ok thanks. I'll give it a shot. It seems like soooo much work to do manual arpabet conversion.

you can't submit pipeline models currently though

I tried again, and so it is.

i'm making a singing voice for talknet should i include lines that have the same lyrics? if no, what about if they're in a different key?

i think i'll include them

if they're at the same pitches they won't help much because talknet is deterministic, meaning it'll always synthesize the exact same audio for the exact same input (with a small context window) different key is very helpful though