What is the Universal model?

jik876 / hifi-gan

HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis

MIT License

1.92k stars 506 forks source link

What is the Universal model? #67

Closed turian closed 3 years ago

turian commented 3 years ago

Could you describe a little more what the Universal model is? It's not in the paper or the audio samples. I'm a little confused what it is.

CookiePPP commented 3 years ago

https://github.com/jik876/hifi-gan/issues/46#issuecomment-757593437 The universal model is just a model trained a very large set of data. More speakers = higher quality with unseen speakers. The goal of the universal model is to have a single vocoder that will work well on almost any speaker without retraining.

turian commented 3 years ago

@CookiePPP Thank you. Do you mind describing what dataset what used for Universal?

jmasterx commented 3 years ago

It's LJSpeech + VCTK + LibreTTS, all downsampled to 22Khz see https://github.com/jik876/hifi-gan/issues/1

turian commented 3 years ago

Thanks, would be great to see that in the README