Codename-RVC-Fork-V4.2.0

update: v4.2.0
| 16.03.2026 |

Aligned up to commit: b6ae29e
This one is a hefty boy lol.

Special thank you to @its5Q for looking into gradio code and fixing quite a few issues!
Directly allowing us to migrate from gradio 5.23.1 → 6.9.0

Changelog;

NEW FEATURES

• CoMBD and SBD discriminators ( They come from Avocodo )

• UnivHD discriminator implem. ( EXP - Trials in APEX-GAN vocoder. )

• f0 / Pitch curve editor added for inference.
  ( Allows for various pitch transformations or corrections. )

UPGRADES

• Increased pitch / harmonic coverage of FGSS in APEX-GAN

• APEX-GAN is now utilizing new [CoMBD, SBD] + UnivHD discriminators.

• Improved tensorboard ( logs ) alignment at resume

• 2-stage grad clipping scheduling added

• Much better and simpler tensorboard opening thanks to @Mikey-Mikey e85411b

CHANGES

• Removed PRLS adversarial loss.

• Corrected the TPRLS adversarial loss. ( Not recommended for general use tho. )

• "ALPEX-GAN" vocoder is renamed to "APEX-GAN" ~ Laplacian and cyc-noise approach proved to be inappropriate for this architecture.

• Non-randomized training mode removed. ( Has no actual purpose )

• BF16 precision is no more supported. ( Not worth it )

• Evaluation mode support removed. It was pretty much useless anyways, mainly due to pesq.

• G/D finetuning is now decoupled. Accepts just one being present as pretrained model.
  ( Generally not needed for casual users. It's mainly for my experiments. )

• Adjusted the hinge loss; made it clamped ( more aligned with how wavehax uses it. )

• Moved heavier/longer ( or those that need fancy formatting ) UI descriptions to external file for ease of editing in future.

• Preprocessing is now supporting these modes for normalization: post_peak ( * 0.95 ), post_peak_rvc ( alpha-blend ), post_rms ( New default. )

• Due to adding f0 editor, I removed support for external f0 curves. ( Because f0 editor supports presets. )

FIXES

• Fixed all UI issues regarding old gradio --> 6.9.0 transition I could find.

• Suppressed ProactorEventLoop connection-reset related error.

• Fix command-line bugs/inconsistencies ~ Thanks to @SirCapybar 353c86d

PERFORMANCE RELATED

• Removed some code-bloat and unused things ( including unused discriminators. )

NOTES
As always, if you notice any bugs or have any issues with any functions,
please report it in issues or on my discord server.

Codename-RVC-Fork-V4.1.0

update: v4.1.0
| 06.02.2026 |

Aligned up to commit: d06aea0
Haha.. this is a pretty huge update if you asked me so, please forgive me if I missed mentioning anything.
remainder for myself to keep on tracking it if I am to take breaks

PS. For this specific update? I just simply recommend to install it all fresh, just in case.

Changelog;

NEW FEATURES

• Added in early-stopping functionality ( can be found in ui).
  Basically lets you pause anytime mid-epoch then resume exactly from that point.

• Added in "pca projection" as an additional metric / feedback.
  ( Reliability still not confirmed, therefore it's not in UI. Only in train script. )

• Rework for preprocessing; Basically, integrated my ml-based 'SmartCutter' into the fork ( UI, preprocessing tab. )
  More info on that can be found in the UI or here
  ( In a short, manual silence-truncation (( hopefully )) is no more needed. Only concatenated-type datasets.

• Tensorboard now automatically handles the over-accumulated steps on init.
  ( Those that are still logged past saved epochs. )

• Heavily experimental "Two-Stage Training Protocol" approach.
  ( Freezes encoders, flow and spk emb + speeds up the lr decay. )
  (( Might help for small datasets but yeah, need more feedback on that. Might get tweaked or removed in future. ))

• Much better "exponential learning rate decay" per-step styled scheduler

• Support for 0-idx_{NAME} style sid-folder naming ( aside of just 0-idx such as "0, 1, 2 .. 108" etc. )
  ( On top of that, includes a safety check against attempts of training with idx-non-contiguous set. )

UPGRADES

• UVMP format now gets an option to have single-for-all index.

CHANGES

• No more "2333333" nonsense. G/D files from now on have their respective step counts included.
  The training always gonna resume from the highest step count ckpt set.
  Everything else remains exactly the same.

• PCPH-GAN undergoes yet another revision of changes Sorry, lied last time about 'final tweak test'

• Removed dead code "init weights" from all vocoders..
  ( It might have had some use back in old pytorch era, not anymore. )

• Got rid of old key remapping for outdated pre-parametrizations weight_norm, this is a nonsense from dated codebases..
  Due to that, I had to remap the pretrains manually.
  They're on hf, will be be re-downloaded automatically if you get rid of old pretrains.. Or simply reinstall the fork from zero.
  ( Unfortunately, from now on, fork officially loses support for Applio-trained pretrains. )
  (( That is, until and if noobies and blaise decide to get rid of archaic code and handlers. ))

• post_lufs and post_lufs_vad options get removed.
  ( I concluded the safest and most reliable approach is per-sample peak norm. )

• Changed some defaults here n' there.

• Enabled weight_decay for AdamW and RAdam ( set to 0.01, aligned with the original RVC behavior. )

• Better gradient logging handling. Now excludes gradscaler skipped steps and infs/nans from main grad graphs.
  ( Those are now logged separately to not skew the averages. )

• kl loss reverted back to it's unclamped version.

FIXES

• Fixed a "ringformer v3" related bug on fp32-only gpus.
  ( Was my mistake to leave that dead code in. )

NOTES
As usual, if unsure of what new features do best to avoid them. Saves time for both you and me.
If you notice any issues or bugs, please report them immediately in fork-and-applio channel ( you can @ me ).
Much appreciated in advance!

Codename-RVC-Fork-V4.0.9

update: v4.0.9
| 30.12.2025 |

Aligned up to commit: 5d99ccd bb98c78
Had to remove 1 line from pcph-gan ^ ( was kept for the sake of beta-test of pretrains. Not used anymore. )

Changelog;

NEW FEATURES

• Custom PCPH-GAN vocoder and so, new architecture. ( In test / prototyping stage. )
• Added envelope loss.
  ( Part of PCPH-GAN arch. but can be enabled from ui independently if you wanna experiment with it. )
• New adversarial losses. ( TPRLS and HINGE but imo best to stick to default / safe LSGAN. )

UPGRADES

• Added seed support for TTS inference.

CHANGES

• Removed support for: Korean, Japanese and Chinese hubert base embedders.
• Removed support for MRF-HiFi-GAN.
• Removed support for snake-hifigan-nsf. ( It evolved into it's successor; PCPH-GAN. )
• Removed support for Wavehax. ( Proved to be too slow and likely unstable in convergence.. )
• Removed periods: 17, 23, 37 for MPD in [MPD, MSD, MRD] combo. ( Used in RingFormer and PCPH-GAN training. )
• Removed bfloat adamw optimizer. ( No point to keep it since for bf16 training, optimi's bf16-friendly adamw is used. )

FIXES

• Fixed TTS inference.

EXTRAS

• Better tensorboard log alignment script in extras.

NOTES
As usual, if unsure of what new features do best to avoid them. Saves time for both you and me.
If you notice any issues or bugs, please report them immediately in fork-and-applio channel ( you can @ me ).

Codename-RVC-Fork-V4.0.8-rev2

update: v4.0.8-rev2
| 03.12.2025 |
Aligned with commit: 6fa5f7c

Changelog;

NEW FEATURES

• Experimental tprls loss ( Disabled for now, it's effectiveness is uncertain as of yet. )
• Linux support ( Installer and run scripts ~ Thanks to 2jang )
• Support for Noobies' newest RefineGan ( Only involved adding his take on mpd+msd+mrd discs. combo )

CHANGES

• Entirely removed zluda support
• SnakeBeta and ISTFT operations in fp32

FIXES

• Add missing 24khz mute for spin v2
• Critical fix for vits1/vits2 encoders import clashing

Codename-RVC-Fork-V4.0.8-rev1-hotfix1

update: v4.0.8-rev1-hotfix1
| 27.10.2025 |

Aligned with commit: 48eb12f

Changelog;

CHANGES

• Changed the projection strength multiplier for AdamSPD from 1.0 to 0.5 ( Seems to be bit safer bet. )

FIXES

• Fixed per-step exp lr decay scheduler when training using bf16 precision.

Codename-RVC-Fork-V4.0.8-rev1

update: v4.0.8-rev1
| 25.10.2025 |

Aligned with commit: c6c7708

Changelog;

CHANGES

• Changed the projection strength multiplier for AdamSPD from 0.001 to 1.0

Codename-RVC-Fork-V4.0.8

update: v4.0.8
| 25.10.2025 |

Aligned with commit: 9de719e

Changelog;

NEW FEATURES

• Added v2 RingFormer configuration files.
  ( Further tests required. )

• Added in Wavehax vocoder. [ Experimental ]
  ( So far no serious tests were done, hence needs trial training to ensure it's stability. )

• Added in snake_hifigan_nsf vocoder. [ Experimental ]
  ( My lil experiment. Has potential to outperform standard nsf-hifigan. )

• Added in AdamSPD optimizer. ( Supposedly tailored for finetuning + uses new weight decay technique. )

• Added in VITS2 architectural upgrades. [ Experimental ]
  ( Transformer-enhanced normalizing flow + spk conditioned text encoder. )

• Added FCPE f0 extraction method
  ( Uses FCPE "DDSP_200k" model variant )

• Added back peak post-normalization method.

• Checkpoint Extractor. Allows to extract small ( usable models ) from big G network .pth files.
  ( Can be found in Utilities -> Model Processing tab. PS. For the time being, only supports HiFf-GAN-NSF models. )

UPGRADES

• RingFormer now supports FP16 training.
  ( Unsure about stability tho. )

• A lil change in KL loss calculation. Swapped to a variant that does not allow negative values ~ clamps to 0.
  ( Not a solution to the problem at all, just a safety-ish feature. )

• Exponential lr decay operates now in two modes: per step and per epoch.
  ( Per step is a new default since it's more fine-tuning friendly and generally more consistent. )

CHANGES

• MRF-HiFi-GAN Added back to the ui.

• HiFiGAN-NSF code now has SineGen included rather than importing.

• Separated flow modules into their dedicated script
  ( To help with organization + make it less messy given I've incorporated some VITS2 elements. )

FIXES

• Corrected triton version for torch 2.7.1 compatibility

NOTES
This is a quite a big update so.. in case something is buggy or broken, please let me know!

Codename-RVC-Fork-V4.0.7-rev1

update: v4.0.7-rev1
| 03.09.2025 |

Aligned with commit: 9214458

Changelog;

UPGRADES

• Support for spin_v2.
• Cyclic Kl loss annealing ( Thanks to @its5Q )

CHANGES

• Removed discs that aren't and won't be used.
• Support for FCPE f0 predictor is removed. No need to use it, esp for training. RMVPE crushes it in almost everything.
• Training code refactoring.
• Peft LoRA experiment unfortunately failed. Not suitable for this arch so I removed it.

FIXES

• Corrected some variable / element names.

QOL

• Training presets ( Thanks to @its5Q )

NOTES:
I advice to delete the "env" folder and re-install the fork.
And as always, if you encounter any issues, please let me know.

Codename-RVC-Fork-V4.0.7

update: v4.0.7
| 06.09.2025 |

Aligned with commit: 61df71a

Changelog;

UPGRADES

• Added in LoRA finetuning option. You can play around it if you want. ( No, such models won't work in applio :> )

CHANGES

• Some imports rearrangement ( Thx to Everyon ).
• A bit refactoring to training code.
• Encoders and many other components are now aligned with mainline ( Better be safe than sorry. Less stress for everyone lol. )

FIXES

• Fixed MPD discriminator ( Strides issue. Same as in Applio. ).

RINGFORMER RELATED

• A bit of code refactoring n code cleanup.
• SineGen and NSF components operate in fp32.
• Configs have their seed changed back to 1234 for the sake infer-preview consistency.

Codename-RVC-Fork-V4.0.6-rev2

update: v4.0.6-rev2
| 25.08.2025 |

Aligned with commit: 75ad9f7

Changelog;

UPGRADES

• Upgraded the target LUFS range finder, making it as robust as it can get.

1. Now It will find the most appropriate target LUFS for your dataset, regardless of what you input for target.
2. While doing the range-test, it does it in 'dry' convention meaning, doesn't save any files until the correct LUFS is found.

• Added in a switch to enable/disable Automatic LUFS range finder.

CHANGES

• Reworked the loudness normalization.

1. Decoupled slicing/resampling from normalization:
   First it does the slicing & resampling, then a dry-run to find the optimal LUFS and lastly the actual norm.
2. The " optimal " found LUFS is based on the worst-slice from all available ( According to its integrated loudness (LUFS) and the sample peak (dBFS) ),
   And so, the true effectiveness is ultimately tied to how good baseline-wise your dataset is, but it's not a big deal.
   In other words, As long your set isn't actually awful, you shouldn't really worry about it too much.

Tl;Dr: Now it's more efficient, quicker and even more robust.
That's most likely the last revision of it. Can't get it any better than that lol.
( PS. Works on huge sets, vctk, multi-speaker etc just fine. )

• MRF-HiFi-GAN completely removed from the ui. From now on, only available via cli.

• Removed "pre" normalization. It's effectively useless. No point keeping it in.

RINGFORMER RELATED

• After some tests and conclusions, the presence of MSD is actually beneficial, hence stays in.