guys, in the ubduck sites some models make on tacotron can sing, do u have any colab, I only see the Talknet version for singing??

tacotron cannot sing

it has zero support for reference audio

what a problem cause i want to make a model in another lenguage with special characters

And talknet use utf-8 only for english, idk

well you can just manually make them sing and pitch correct and time correct it in post

very hard though

if you mean that they have

(singing)

in the name, that just means they're trained on singing audio; they can't use reference audio like talknet can if you mean the pitch and pace controls that like, 4 rappers have so far, that's a separate model called fastpitch that was tested a bit (zwf and co are working on something even better though)

talknet actually doesn't work with any language other than english even if the characters are ascii only, because it converts the messy english spelling into phonemes (

l ay k dh ih s

) it's theoretically possible to make it a different language but would require serious work

I'm developing a TTS website using Flask and I was testing with some friends and sometimes the audio that was generated to my friend was passed to me and vice versa, do you guys know any solution to that?

I was thinking about making it so the generated file would have a random name every time, but I don't think that would work too

uuid can help

many websites do exactly that, including 15.ai

personally i have no idea how webdev works but 15 just takes a few seconds before it responds to the request with the url and all

Thinking about, I'll try using uuid and making a python code in the flask part to erase the audios from the folder everyday

I thought that it wouldn't work because it could mess up sending the files if two persons generated an audio at the same time

But probably it will work fine

i thought the point of the uuid thing is that it would prevent that exact mess-up?

Yes, but I don't know exactly how Flask sends the audio to the person

I don't know, In my head I thought about the possibility of flask sending the wrong audio to another person

Probably won't happen

It shouldn't happen

Flask does stuff synchronously

sooooo i trained a TalkNet model.. and this is what it sounds like when I use the controllable notebook to generate audio without a reference input "Ever on and on, I continue circling with nothing but my hate and the carousel of agony" (lyrics of bad apple translated to english)

i'm pretty sure this has to do with my training data but why does it talk so fast 😰

100% the reference audio

wdym

i was generating without reference audio

or do you mean its my training data being the problem

oh it’s without reference audio