I am blown away by the accuracy of the so-vits 5 fork, and it's ability to to get trained models the accuracy of complex lyrics, like Rap God by Enimem!
https://github.com/PlayVoice/so-vits-svc-5.0/tree/bigvgan-mix-v2
I've attached a wav file of a no-pitch sample export I did.
(no pitch means wubby sings mostly monotone, as I need to fix inference)
removing old model
NOTE: FINAL UPLOAD OF MODEL USING SO-VITS-FORK 4.0
New model being generated with so-vits-fork 5.0, with separate pitch and hubert inference :)
removing old model
NOTE: FINAL UPLOAD OF MODEL USING SO-VITS-FORK 4.0
New model being generated with so-vits-fork 5.0, with separate pitch and hubert inference :)
This is a clip of the song Boom Boom by Venga Boys, that I clipped and vocal-removed, as well as BGM-removed, for you to test the model with.
If you're training it yourself, after you gracefully stop the train, you can test with 'svc infer sample.wav'
The source wav files I used to build this model from.
Use if you want to train your own model, or if you want to submit even further optimized wav files for a new model train.
250 Epochs as of upload date - will upload as updated/trained.
Trained using a Wubby Sleep Aid video that I removed BGM, isolated vocals, trimmed to 9sec clips and then loaded it into so_vits_svc fork. 5 hours of perfectly trimmed audio for our regards :)
Trained on (two rigs):
Initial 24hr train, first 100 epochs: i7 11700k, 3070ti 8gb, 32gb DDR5.
All training after: i9 13900k, 2x 4090ti 48gb, 128gb DDR5
Sample audio at 250 attached.
