Skip to content

Latest commit

 

History

History
90 lines (77 loc) · 4.74 KB

README.md

File metadata and controls

90 lines (77 loc) · 4.74 KB
Raw

RVC Inference

English | 中文简体 | 日本語 | 한국어 | Français| Türkçe

Translations provided by GPT-4.

This project is a lightweight, fast, and memory efficient api that runs v1/v2 RVC models. It is intended for use in production environments and compatibility with existing codebases. It makes integrating RVC as a stage in a pipeline or workflow easy. Installation is quick using pip and should be compatible with Linux/Windows/Mac and the latest python versions.

Install

If using Python 3.11+ install the fairseq fork first as fairseq is not yet compatible with 3.11. (Will take a minute).

pip install https://github.com/One-sixth/fairseq/archive/main.zip

Pip install the repo like below and all dependencies will be installed automatically.

pip uninstall inferrvc
pip install https://github.com/CircuitCM/RVC-inference/raw/main/dist/inferrvc-1.0-py3-none-any.whl --no-cache-dir

By default pypi installs the pytorch cpu build. To install for gpu using Nvidia or AMD, visit https://pytorch.org/get-started/locally/ and pip install torch and torchaudio with gpu before installing this library.

Support should be available for Python 3.8-3.12 but only 3.11 was tested. If there are any problems with installation or compatibility please open an issue and I'll push out fixes. PR's with fixes and improvements are welcome.

Usage

First set the optional environment variables:

import os
os.environ['RVC_MODELDIR']='path/to/rvc_model_dir' #where model.pth files are stored.
os.environ['RVC_INDEXDIR']='path/to/rvc_index_dir' #where model.index files are stored.
#the audio output frequency, default is 44100.
os.environ['RVC_OUTPUTFREQ']='44100'
#If the output audio tensor should block until fully loaded, this can be ignored. But if you want to run in a larger torch pipeline, setting to False will improve performance a little.
os.environ['RVC_RETURNBLOCKING']='True'

Notes on environment variables:

  • Both RVC_OUTPUTFREQ and RVC_RETURNBLOCKING set defaults for the RVC class, but they can be overriden per instance with self.outputfreq and self.returnblocking.
  • Setting RVC_OUTPUTFREQ to None will disable standard resampling and return the model's native sample rate.
  • If you do not set RVC_INDEXDIR the RVC class will fallback to RVC_MODELDIR and lastly the absolute path of the model directory os.path.dirname(model_path).
  • If you do not set RVC_MODELDIR then arg model must be an absolute path.

Load models:

from inferrvc import RVC
whis,obama=RVC('Whis.pth',index='added_IVF1972_Flat_nprobe_1_Whis_v2'),RVC(model='obama')

print(whis.name)
print('Paths',whis.model_path,whis.index_path)
print(obama.name)
print('Paths',obama.model_path,obama.index_path)
Model: Whis, Index: added_IVF1972_Flat_nprobe_1_Whis_v2
Paths Z:\Models\RVC\Models\Whis.pth Z:\Models\RVC\Indexes\added_IVF1972_Flat_nprobe_1_Whis_v2.index
Model: obama, Index: obama
Paths Z:\Models\RVC\Models\obama.pth Z:\Models\RVC\Indexes\obama.index

Run Inferencing:

from inferrvc import load_torchaudio
aud,sr = load_torchaudio('path/to/audio.wav')

paudio1=whis(aud,f0_up_key=6,output_device='cpu',output_volume=RVC.MATCH_ORIGINAL,index_rate=.75)
paudio2=obama(aud,5,output_device='cpu',output_volume=RVC.MATCH_ORIGINAL,index_rate=.9)

import soundfile as sf

sf.write('path/to/audio_whis.wav',paudio1,44100)
sf.write('path/to/audio_obama.wav',paudio2,44100)

Whis example.
Obama example.

Changes from the original repo:

  • Removed most code not related to inferencing. Now much fewer dependencies.
  • Made a streamlined inference class and pipeline.
  • Performance and memory efficiency improvements.
  • Generic models are now managed by huggingface_hub and cached through the path HF_HOME environment variable.
  • Flexible referencing of RVC model directory and files.
  • Disabled the butterworth filter by default as there is usually no difference and might slightly reduce quality. Can be enabled with inferrvc.pipeline.enable_butterfilter=True.

Todos:

  • Test different python versions.
  • Test different OS's and pitch estimators. (The other estimators should be ported but only RMVPE was tested, it is the best)
  • Move the remaining operations to the single primary device (eg gpu), to reduce latency and slow down resulting from memory transfers.
    • Replace remaining numpy code with torch equivalents torch.where and torch.masked_select.
    • Reimplement the index mask with pytorch for gpu devices.
  • Make use of torch 2.0 .compile() to speed up the v1/v2 models if possible.