rvc-model
The model files used by rvc-next, a rewrite of Retrieval-based-Voice-Conversion-WebUI (RVC), organised by category, together with the official RVC demo voices.
Everything here is copied unchanged from the official RVC repository,
lj1995/VoiceConversionWebUI (revision
e6d0c1a), except the FCPE and CREPE pitch models, copied unchanged from the
torchfcpe 0.0.4 and torchcrepe
0.0.24 packages, which bundle them. The demo voices come
from inside RVC's distribution packages (RVC*.7z), where they are the only copies.
manifest.json lists every file with its size, SHA-256 and where it came from.
Contents
| Folder | Files | Used for | Upstream path |
|---|---|---|---|
hubert/hubert_base/ |
config.json, preprocessor_config.json, pytorch_model.bin |
Content features (HuBERT/ContentVec), for conversion and training | hubert_base/ |
rmvpe/ |
rmvpe.pt, rmvpe.onnx |
Pitch extraction (RMVPE); the ONNX export for DirectML | rmvpe.pt, rmvpe.onnx |
fcpe/ |
fcpe_c_v001.pt, LICENSE-FCPE.txt |
Pitch extraction (FCPE) | torchfcpe/assets/fcpe_c_v001.pt in torchfcpe 0.0.4 |
crepe/ |
full.pth, tiny.pth, LICENSE-CREPE.txt |
Pitch extraction (CREPE full and tiny) | torchcrepe/assets/ in torchcrepe 0.0.24 |
pretrained/v1/ |
{f0,}{G,D}{32k,40k,48k}.pth |
Base models for training v1 voices, with (f0) and without pitch guidance |
pretrained/ |
pretrained/v2/ |
the same twelve files | Base models for training v2 voices | pretrained_v2/ |
separation/ |
five BS-/Mel-Roformer checkpoints and their YAML | Vocal separation, dereverb and karaoke (PyMSS) | pymss_weights/ |
voices/<name>/ |
<name>.pth, <name>.index |
The official demo voices | inside the RVC*.7z packages |
Demo voices
| Voice | Version | Rate | Pitch guidance | Index |
|---|---|---|---|---|
kikiV1 |
v1 | 40k | yes | yes |
keruanV1 |
v1 | 40k | yes | yes |
guanguanV1 |
v1 | 40k | yes | yes |
zhanzhanv2-xi |
v2 | 40k | yes | yes |
youzhanv2-xi |
v2 | 48k | yes | no (none was ever published) |
Each is an RVC small model (.pth) with its retrieval index (.index); they load in the original
RVC WebUI as well as in rvc-next (rvc-next model download --all, or Models › Voices › Demo
voices).
Licence
MIT, as the upstream repository, including its authors' condition (in LICENSE-RVC.txt): the
software is for research use only, and users bear full responsibility for the voices they produce
and distribute; anyone who does not accept this may not use these files. The licences of the
components the models build on (ContentVec, VITS, HiFi-GAN, UVR, audio-slicer) are listed in the
same file.
The FCPE model (fcpe/) is not part of RVC: it is MIT-licensed by its authors (Copyright (c) 2023
CN_ChiTu), without the research-use clause; its licence is fcpe/LICENSE-FCPE.txt. The CREPE
models (crepe/) are not part of RVC either: MIT-licensed by torchcrepe's author (Copyright (c) 2020
Max Morrison) and converted from CREPE (Copyright (c) 2018 Jong Wook Kim), without the research-use
clause; see crepe/LICENSE-CREPE.txt.