RVC model very 'glitchy', I guess

[barren citrus]

Hello, does anyone know why does my trained model sound so 'blurry'? It's not clear at all, even if I trained it with 40 mins of clean audio of me reading some texts. The model is trained for 200 epochs and every time I try to generate some song cover, my generated voice is so... strange?! You can hear in the audio file attached.

[barren citrus]

bump

[surreal flint]

Is your dataset just talking

[barren citrus]

yes, it is. should i also sing? if i want to generate cover songs with my cloned voice
i mean, the answer probably is obviously yes :)) but i thought somehow the RVC can handle this situation of giving only reading text as input

because I see tons of fake videos with politicians singing, and I'm sure who created their cloned voices gave as input only some vocals with those politicians talking, or from interviews or things like that, so no way they had recordings with that public people singing :))

so that's why i thought: if they can do this only with 5 mins of random words, then I can also do with 45 full minutes of words..
more than that, the words are lyrics of the songs i want to clone, so they're not just random words.. but yes, the problem may be the fact that i just read them, instead of singing them

can you confirm?

[lusty helm]

chances of doing it with just 5 minutes of non-rich in pitch variations speech are incredibly small

Generally, getting a good singing-capable model out of speech ( again, with rather monotone traits )
is quite difficult

Try smaller batch size and / or use KLM pretrains instead of stock / original ones.

[proven oxide]

better pitch range/expressiveness, better capability to sing