SeoulStreamingStation commited on
Commit
9b13040
·
verified ·
1 Parent(s): 9b768b5

Upload 2 files

Browse files

/////////////////////////////////////////////////////////////////////////////////
RVC KLM9 Pretrained Model - Hifigan / ContectVec / Dev Test Model
/////////////////////////////////////////////////////////////////////////////////

KLM9 replaces the Spin2 embedder with the original ContentVec.

It was pre-trained using BF16 with a batch size of 128 (32×4) on a large-scale voice dataset consisting of 4,260 hours of audio and 2,635 speakers, including high-, medium-, and low-quality recordings (such as telephone-quality audio).

After that, a second fine-tuning stage was performed using 550 hours of high-quality audio data from 109 speakers.

Compared to previous KLM series models, it reproduces relatively stronger sibilance and non-vocal sound regions more effectively.

HIFI-GAN
Pitch Extraction - RMVPE
Embedder - ContentVec
AdamW
BF16

Files changed (2) hide show
  1. D_KLM9T_HFG_32k.pth +3 -0
  2. G_KLM9T_HFG_32k.pth +3 -0
D_KLM9T_HFG_32k.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:56fd4f1782119a80c00ff11abb46f263d278d1e2b8833a1bdfa3be23fbe94fa9
3
+ size 857123185
G_KLM9T_HFG_32k.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:82b8bb37ac21d13aa0ceb2baff3d9159983834b04369c3a97d1da78f052e525c
3
+ size 443229213