Nonattentive tacotron will use ref audio It's in production rn

I heard about that I am really excited for it

I believe I never posted any other NNSVS synthesis examples here of mine

I'm mainly working on making Polish support for it, as well as some extra resources to make it easier to train a model for this language

Wow. Amazing. I have so many questions. Did you get your 330 samples from multiple acapellas? Is that alright because if someone is singing in different styles they shouldn’t be in the same dataset, right? I also find myself skipping lots of parts of a song that have background singing. Did you clean up the headphone bleed on the acapellas before training? Or were the acapellas already perfectly isolated vocals? Approximately how many seconds per sample do you recommend?

This was them actually voicing this bank

I’ll look into this. Does nnsvs have the option for reference audio?

No

You'll need to tune and make a midi

i think you could patch it together with like

there was a vocalistener-like plugin for utau

but mostly yeah midi is the way to go

Oh yeah something like that would be. Sorta similar.

Interesting OK

btw how long does it take to label 1 minute of audio

Oh that highly depends on the person

how long does it take for you then

I've never timed I can later

for me id take like idk 3 mins?

if im transcribing too

then maybe

10 mins

It's been a while since I've done much with talknet, I basically just sang a bunch of songs I knew the best by heart and chopped them up where I felt it was appropriate. The samples were recorded in my studio on my RE20 so they were pretty damn clean but I slightly gated the hum of my fridge. For me I just dumped everything in the same dataset but I recommend using things that stay relatively consistent.

or something

I'm probs gonna make 3 versions of my talknet bank Soft, normal, and power I'll use the data i'll use for eng nnsvs but segment it

Cuz for nnsvs i'm gonna do soft/normal/power and flag the appends

That sounds fun I haven't decided how I am gonna segment my next recording sesh because I am unsure how singing all these songs that were giving to me is gonna go 😂

Love how everyone just ignored that

I didnt ignore it in my head

Didnt reply tho, sorry