changing the vocoder

[HobisPL]

Have you tried changing the vocoder from Waveglow to HiFi-GAN? HiFi-GAN is faster and requires less VRAM. Alternatively, you could try adding a different vocoder.

[152334H]

Note that TorToiSe uses Unvinet, not Waveglow. I've received the following other suggestions on top of HiFi-GAN:

• BigVGAN, which was just released and allegedly is better than HiFi-GAN universally
• AdaVocoder, for better fine-tuning
• some other ones i forgot about

all good options

[deviandice]

@152334H got bored of waiting so learned python and did it myself out of boredom, now it sounds way less shit :D https://github.com/deviandice/tortoise-tts-BigVGAN Make of it what you will. Here's the model. https://disk.yandex.com/d/fOjzTs8HQiFVdg

Wouldn't mind some credit if at all possible, thanks.

[Ryu1845]

See also mrq's adaptation for more inspiration https://git.ecker.tech/mrq/ai-voice-cloning/issues/52

[deviandice]

See also mrq's adaptation

That's literally also my implementation haha

[152334H]

ok i added bigvgan but i am still not going to change much in this repo in the future

you should also tell that mrq fella to use submodules/packages more often

---

fwiw @deviandice, I copied none of the code from your impl, but either way I added some attribution to the README.

[frandmb]

Is there any benchmark or estimate using Bigvgan against the previous vocoder?

[Ryu1845]

You can look at the figures of the paper https://www.semanticscholar.org/paper/BigVGAN%3A-A-Universal-Neural-Vocoder-with-Training-Lee-Ping/04f5553934c458305a501d63323f1b841fd5d102

[deviandice]

fwiw @deviandice, I copied none of the code from your impl, but either way I added some attribution to the README.

That means a lot, thank you. It's not every day someone cites my work like this.