#(✅ SOLVED) RVC Training Help!

1 messages · Page 1 of 1 (latest)

signal trout
#

I'm a beginner in RVC training, and I'm trying to figure out what I'm doing wrong.

The issue: When training a voice model with the 42 minute dataset I created as shown in the image provided, the results are really really bad. It can't even form words, and when I load in other models created by other people, their models are noticeably way way better than my voice model.

Computer Specs:
GPU: RTX 5070 TI
CPU: I-7 14700k
RAM: 32 GB DDR5

App Used For Training: Applio V3.6.0
Pretrain: Legacy Core V1.5

(I'm unable to upload my Dataset audio since the file is too big to attach)
(I truncated the silence and normalized the audio with Audacity)
(Also, I did try the default Applio settings for preprocessing the dataset, and it was still just as bad as my custom settings)

4iojwI8gDcRp4cW0Dwa7.png UrdODdiGp91ztwCSsEue.png aLj65jru4z9fVQl2OA3G.png wCo1ruuGfwh5ckXzJcAz.png bKkzurKzbvixPj1aerQ6.png MpdYUmWYNXbohvqBHFSo.png gEUx5cJOCLcs9Iy64Nrf.png VFAhflQvEILj1iXj2EUT.png
light nebula
#

w-okada can only inference contentvec models

use tg-develop's w-okada fork or vonovox to inference spin models

#

i recommend vonovox because is faster

#

and it was made by the guy who created spin for rvc

signal trout
#

Ah okay, I'll try that right now and lyk if it resolves my issue

#

But where can I download it?

light nebula
#

download VONOVOX_1_7_BETA.zip

#

unzip it

#

then download vonovox_beta7_fix6.zip

#

unzip in the same folder you have vonovox 1.7 beta, and replace the files

#

and you can start the app by running the start.bat file (srry for the bad english, its not my first language yt_nails )

signal trout
#

Okay, thanks!

signal trout
signal trout
light nebula
#

check if you dont have a filter named "low quality mic" enabled in vonovox

#

if that setting is off, try training a contentvec model instead

#

and see if it fixes that

signal trout
#

Alright

signal trout
# light nebula and see if it fixes that

Thanks! The problem was the spin-v2 embedder model. But that has me wondering: Why does training with spin-v2 sound this bad? How can I adjust it to where the spin-v2 embedder model works without any low quality inference?

light nebula
#

no idea why spinv2 is like that with hifigan

signal trout
#

Hmmm, okay. Is it like that with the version of spin before v2? spin-v1 perhaps if that's what its called?

light nebula
#

it was even more muffled

signal trout
#

Oh damn, I was not expecting that lmao

#

Okay, well your help solved my issue

#

Thank you for helping!

light nebula