- Undo SQL change for the sake of stability, uses csv now
- Merging checkpoints bug has been resolved
- Fixed the cli inferencing traceback.
- SQL Databases implemented for formanting training stop.
- Gradio browser tab renamed to
Mangio-RVC-Web 💻. - Rudimentary functions from
infer-web.pyremoved. - Formanting now accepts any audio format, as long as it is supported by FFmpeg.
- Better MacOS installation script. For inference, all that needs to be done is running the
run.shfrom the extracted zip folder, where it will install Python 3.8, Homebrew, and other dependencies for you automatically. M1 Macs are natively supported for GPU acceleration, and training should work if you choose to download the pretrained models.
- Fp16 detection now works how it did before the last RVC beta; no more training slowdowns compared to how it was before then
- Logging on TensorBoard is now done per epoch. It logs every X steps in an epoch; therefore each point on the graph is now represented per-epoch.
- Updated UI to include placeholders.
- Certain sliders now appear only when needed.
- Step 2b (feature extract) is hidden if pitch guidance is disabled.
- Unnecessary radios have been replaced with checkboxes.
- Experimental Formant Shift using StftPitchShift(tried using praat with praatio but to no avail)
- Added
Stop Trainingbutton when training, no need to restart RVC every time you want to stop the training of a model! - Auto-detect Index path for models selected + Auto-detect paths, no more default values like this:
E:\codes\py39\vits_vc_gpu_train\logs\mi-test-1key\total_fea.npy, We're getting Root Dir and subfolders using
os.path.abspath(os.getcwd())- Audio files dropdown by autodetecting files in
/audios/folder - More stable version of Gradio(3.34.0) with support of Themes
- Removed
One Click Trainingbutton inTrainingTab, since it was glitchy and has confused a lot of users. - Changed default training settings to be more optimal for newer users.
- Auto-open TensorBoard localhost URL when
tensor-launch.pyis executed - RMVPE implemented in both inferencing and training
Im developing alone. Help support me in the rapid development of open-source A.I Deep NN audio tools and frameworks. By donating, you making a commitment to the improvement of audio A.I as a whole. You should donate if you merely want to "support" me, not because you are then expecting a finished product. There are no guarantees. Thank you
💓 Please support the original RVC repository. Without it, obviously this fork wouldn't have been possible. The Mangio-RVC-Fork aims to essentially enhance the features that the original RVC repo has in my own way. Please note that this fork is NOT STABLE and was forked with the intention of experimentation. Do not use this Fork thinking it is a "better" version of the original repo. Think of it more like another "version" of the original repo. Please note that this doesn't have a google colab. If you want to use google colab, go to the original repository. This fork is intended to be used with paperspace and local machines for now.
Now supports version 2 pre-trained models!
I am able to communicate with you here and there.
Special thanks to discord user @kalomaze#2983 for creating a temporary colab notebook for this fork for the time being. Eventually, an official, more stable notebook will be included with this fork. Please use paperspace instead if you can as it is much more stable.
The original RVC Demo Video here!
Realtime Voice Conversion Software using RVC : w-okada/voice-changer
The dataset for the pre-training model uses nearly 50 hours of high quality VCTK open source dataset.
High quality licensed song datasets will be added to training-set one after another for your use, without worrying about copyright infringement.
- f0 Inference algorithm overhaul: 🌟
- Added pyworld dio f0 method.
- Added another way of calculating crepe f0 method. (mangio-crepe)
- Added torchcrepe crepe-tiny model. (Faster on inference, but probably worse quality than normal full crepe)
- Modifiable crepe_hop_length for the crepe algorithm via the web_gui and CLI.
- f0 Crepe Pitch Extraction for training. 🌟 (EXPERIMENTAL) Works on paperspace machines but not local mac/windows machines. Potential memory leak. Watch out.
- Paperspace integration 🌟
- Paperspace argument on infer-web.py (--paperspace) that shares a gradio link
- Make file for paperspace users
- Tensorboard access via Makefile (make tensorboard)
- Total epoch slider for the training now limited to 10,000 not just 1000.
- Added CLI functionality
- added --is_cli flag on infer-web.py to use the CLI system.
- f0 hybrid (median) estimation method by calculating nanmedian for a specified array of f0 methods to get the best of all worlds for all specified f0 methods. Only for CLI right now. Soon to be implemented into GUI 🌟
- f0 hybrid (median) estimation method on f0 feature extraction (training). (VERY EXPERIMENTAL PROBABLY EXTREMELY BUGGY). Feature extraction with the hybrid method will take MUCH longer.
- Reduce tone leakage by replacing source feature to training-set feature using top1 retrieval;
- Easy and fast training, even on relatively poor graphics cards;
- Training with a small amount of data also obtains relatively good results (>=10min low noise speech recommended);
- Supporting model fusion to change timbres (using ckpt processing tab->ckpt merge);
- Easy-to-use Webui interface;
- Use the UVR5 model to quickly separate vocals and instruments.
- An inference batcher script. Improvement Proposal:
- According to various github users, apparently inferencing 30 second samples at a time both increases quality and prevents harvest memory errors.
- Improved GUI (More convenience).
- Automatic removal of old generations to save space.
- More Optimized training on paperspace machines
Right now, the hybrid f0 method is only available on CLI, not GUI yet. But basically during inference, we can specify an array of f0 methods, E.G ["pm", "harvest", "crepe"], get f0 calculations of all of them and 'combine' them with nanmedian to get a hybrid f0 signal to get the 'best of all worlds' of the f0 methods provided.
Here's how we would infer with the hybrid f0 method in cli:
MyModel.pth saudio/Source.wav Output.wav logs/mi-test/added.index 0 -2 hybrid[pm+crepe] 128 3 0 1 0.95 0.33Notice that the method is "hybrid[pm+crepe]" instead of a singular method like "harvest".
hybrid[pm+harvest+crepe]
# the crepe calculation will be at the 'end' of the computational stack.
# the parselmouth calculation will be at the 'start' of the computational stack.
# the 'hybrid' method will calculate the nanmedian of both pm, harvest and crepeMany f0 methods may be used. But are to be split with a delimiter of the '+' character. Keep in mind that inference will take much longer as we are calculating f0 X more times.
The original repos crepe f0 computation method is slightly different to mine. Its arguable that in some areas, my crepe implementation sounds more stable in some parts. However, the orginal repo's crepe implementation gets rid of noise and artifacts much better. In this fork, my own crepe implementation (mangio-crepe) uses a customizable crepe_hop_length feature on both the GUI and the CLI which the original crepe doesnt have. Please let it be known, that each implementation sounds slightly different, and there isn't a clear "better" or "worse". It all depends on the context!
If one must be chosen, I highly recommend using the original crepe implementation (not this fork's) as the developers of RVC have more control on fixing issues than I have.
Crepe training is still incredibly instable and there's been report of a memory leak. This will be fixed in the future, however it works quite well on paperspace machines. Please note that crepe training adds a little bit of difference against a harvest trained model. Crepe sounds clearer on some parts, but sounds more robotic on some parts too. Both I would say are equally good to train with, but I still think crepe on INFERENCE is not only quicker, but more pitch stable (especially with vocal layers). Right now, its quite stable to train with a harvest model and infer it with crepe. If you are training with crepe however (f0 feature extraction), please make sure your datasets are as dry as possible to reduce artifacts and unwanted harmonics as I assume the crepe pitch estimation latches on to reverb more.
Only for CLI (not implemented in GUI yet). Basically the same as usage described in this readme's f0 hybrid on inference section. Instead of stating "harvest" into your arguments in the f0 feature extraction page, you would use "hybrid[harvest+dio+pm+crepe]" for example. This f0 nanmedian hybrid method will take very long during feature extraction. Please, if you're willing to use hybrid f0, be patient.
This is due to the number of processes (n_p) being too high. Make sure to cut the number of threads down. Please lower the value of the "Number of CPU Threads to use" slider on the feature extraction GUI.
Welcome to RVC version 2!
Please note that version 2 pre-trained models only support 40k model sample rates. If you want to use 32k or 48k however, please use version 1 pre-trained models.
Using pip (python3.9.8 is stable with this fork)
cd Mangio-RVC-Fork
make install # Do this everytime you start your paperspace machineNotice: faiss 1.7.2 will raise Segmentation Fault: 11 under MacOS, please use pip install faiss-cpu==1.7.0 if you use pip to install it manually. Swig can be installed via brew under MacOS
brew install swigInstall requirements:
pip install -r requirements.txtIf you're experiencing httpx invalid port errors please insteall httpx==0.23.0
cd Mangio-RVC-Fork
# Do only once after cloning this fork (No need to do it again unless pre-models change on hugging face)
make basev1
# or if using version 2 pre-trained models.
make basev2RVC requires other pre-models to infer and train. You need to download them from our Huggingface space.
Here's a list of Pre-models and other files that RVC needs:
hubert_base.pt
./pretrained
./uvr5_weights
If you want to test the v2 version model (the v2 version model changes the feature from the 256-dimensional input of 9-layer hubert+final_proj to the 768-dimensional input of 12-layer hubert, and adds 3 cycle discriminators), an additional download is required
./pretrained_v2
#If you are using Windows, you may need this file, skip if ffmpeg and ffprobe are installed; ubuntu/debian users can install these two libraries through apt install ffmpeg
./ffmpeg
./ffprobecd Mangio-RVC-Fork
make run-uiThen click the gradio link it provides.
# use --paperspace or --colab if on cloud system
python infer-web.py --pycmd python --port 3000cd Mangio-RVC-Fork
make run-clipython infer-web.py --pycmd python --is_cliMangio-RVC-Fork v2 CLI App!
Welcome to the CLI version of RVC. Please read the documentation on https://github.com/Mangio621/Mangio-RVC-Fork (README.MD) to understand how to use this app.
You are currently in 'HOME':
go home : Takes you back to home with a navigation list.
go infer : Takes you to inference command execution.
go pre-process : Takes you to training step.1) pre-process command execution.
go extract-feature : Takes you to training step.2) extract-feature command execution.
go train : Takes you to training step.3) being or continue training command execution.
go train-feature : Takes you to the train feature index command execution.
go extract-model : Takes you to the extract small model command execution.
HOME:Typing 'go infer' for example will take you to the infer page where you can then enter in your arguments that you wish to use for that specific page. For example typing 'go infer' will take you here:
HOME: go infer
You are currently in 'INFER':
arg 1) model name with .pth in ./weights: mi-test.pth
arg 2) source audio path: myFolder\MySource.wav
arg 3) output file name to be placed in './audio-outputs': MyTest.wav
arg 4) feature index file path: logs/mi-test/added_IVF3042_Flat_nprobe_1.index
arg 5) speaker id: 0
arg 6) transposition: 0
arg 7) f0 method: harvest (pm, harvest, crepe, crepe-tiny)
arg 8) crepe hop length: 160
arg 9) harvest median filter radius: 3 (0-7)
arg 10) post resample rate: 0
arg 11) mix volume envelope: 1
arg 12) feature index ratio: 0.78 (0-1)
arg 13) Voiceless Consonant Protection (Less Artifact): 0.33 (Smaller number = more protection. 0.50 means Dont Use.)
Example: mi-test.pth saudio/Sidney.wav myTest.wav logs/mi-test/added_index.index 0 -2 harvest 160 3 0 1 0.95 0.33
INFER: <INSERT ARGUMENTS HERE OR COPY AND PASTE THE EXAMPLE>cd Mangio-RVC-Fork
make tensorboardThen click the tensorboard link it provides and refresh the data.
-
[OPTIONAL] Change Gradio's theme:
- Open
infer-web.pyin any code/text editing software (e.g.notepad++,notepad,vscode, etc)
2a. Press Ctrl+F and search for
with gr.Blocks(, select the one that's not fully commented2b. Go to line
1842, you'll see thewith gr.Blocks(theme='HaleyCH/HaleyCH_Theme') as app:- Go to Gradio Theme Gallery:
3.1 Select any theme you like (e.g. this one)
3.2 Look at the top of the page
3.3 Copy theme variable(in this case, it's
theme='freddyaboulton/dracula_revamped')- Replace
theme='HaleyCH/HaleyCH_Theme'ininfer-web.pywith any value of a theme from Gradio Theme Gallery
- Open
If you are using Windows, you can download and extract RVC-beta.7z to use RVC directly and use go-web.bat to start Webui.
There's also a tutorial on RVC in Chinese and you can check it out if needed.