btw, just coming out of nowhere with this lol, has anyone ever tried getting tacotron2 to work with the WORLD vocoder? it's kinda pretty prominent in the eastern TTS genre, and I was hoping to get it to work on a personal tacotron2 project I'm working on lol

Does Tacotron2 run decently on CPU? Or it needs to be a GPU.

Training would probably take a long time, but synthesis would probably be okay, depends on CPU of course

github link?

Here yo go, thanks for lookin into it. Hifi Gan and Talknet are not bad either, but World is like pretty quick at synthesis too imo

Japanese and Chinese models sounds more realistic than English model

i don't speak those languages, so they sound great.

i think its similar to VQGan (taming transformers)-- where when you try it on a bird nothing seems out of the ordinary, but on a face we know the problem and find it unrealistic

speaking of vqgan

LOL awesome-- share it in the dall-e discord server

also this pretty cool

thats kinda a given since those languages only have 1 way of pronouncing stuff rather than english's 02382832329 ways

Imagine using CLIP with it and then make it sound like Rick Astley

Gotta admit tho, the JSUT tts voice for ESPnet2-TTS is hella impressive compared to the LJSpeech tts voice lol, after all the latter has about 24 hours of material

also nice to see u here too

this seems interesting

nice, posted the paper for it above

is there a trend for singing voice synthesis papers there is rarely code published?

have a space for it here now https://huggingface.co/spaces/akhaliq/SpecVQGAN_Neural_Audio_Codec

one issue with the example.wav file

i think it should work now after this build

This is well into the 70s.

I have some good news and bad news for my voice of DECTalk (Used in moonbase alpha)