A modernized, high-performance RVC fork of w-okada/voice-changer optimized for pure ONNX Runtime and modern CUDA environments.
This project is a modernized spiritual successor to w-okada/voice-changer (a.k.a. W-Okada VC), originally created by w-okada. We deeply respect and honor w-okada's pioneering contributions to the real-time AI voice conversion community.
Note
For Japanese speakers: 日本語で読まれる場合は、ブラウザの翻訳機能（右クリック「日本語に翻訳」など）をご利用ください。
Download the standalone package matching your hardware from GitHub Releases:
If you have an NVIDIA GPU, download the package with the cuda edition for optimal performance; otherwise, the cpu edition is recommended.
Extract the archive to any folder and run:
- Desktop App (Recommended):
Double-click
vcclient-forge.exe(Windows) or execute./vcclient-forge(Linux / macOS). - Web Browser Mode (Classic):
Double-click
start.bat(Windows) or execute./start.sh(Linux / macOS). Your default web browser will automatically open:http://localhost:18888/
Note
Initial model assets will be downloaded automatically on first launch. This may take a few moments depending on your internet connection.
To transfer your existing models and settings, simply copy the settings and model_dir folders from your original v2.0.78 directory into the extracted Forge folder.
- OS: Ubuntu 24.04 LTS or newer (
glibc >= 2.39). - NVIDIA Driver (
cudaedition only): Host NVIDIA GPU driver is required (CUDA Toolkit and cuDNN are bundled and not required). - Audio, Desktop & Archive Dependencies:
# Ubuntu / Debian sudo apt-get update && sudo apt-get install -y libportaudio2 libwebkitgtk-6.0-4 7zip
- Spiritual Successor to the Golden Standard (v2.0.78):
Designed as a community-driven spiritual successor to the most widely adopted and beloved version (
v2.0.78), honoring the original vision while breathing new life into the established ecosystem. - Windows-First Portable Experience: Prioritizing the standalone Windows ZIP distribution as our first-class citizen. Our primary mission is ensuring seamless, zero-config local execution on Windows gaming and streaming rigs.
- Seamless Usability & Compatibility (Zero-Config Transition): Preserving the familiar UI workflows, model slots, and configuration compatibility of v2.0.78. Creators and streamers can transition immediately without re-learning or re-configuring their setups.
- Laser-Focused on Real-Time Performance: Dedicated strictly to real-time voice conversion—maximizing latency efficiency and audio stability under demanding gaming and streaming workloads.
- Dedicated RVC Client: Exclusively specialized for RVC to deliver a clean, efficient, and intuitive real-time voice conversion experience.
- Future-Proof Maintainability: Engineered to smoothly track modern Python (3.14+) and newer GPU driver architectures, ensuring long-term reliability without bit-rot.
- Decoupling PyTorch toward a "Pure-ONNX" Runtime: Aiming to streamline the entire inference pipeline exclusively into ONNX. By intentionally shedding heavy PyTorch dependencies, we dramatically minimize dependency conflicts and shrink standalone distributions to a fraction of their original size.
- Deterministic Package Management (uv):
Employing
pyproject.tomland a locked dependency manifest (uv.lock) for lightning-fast, 100% reproducible environments. - Platform Scope & Support Boundaries:
- Windows: Primary target. Officially supported and validated for portable ZIP distributions and source execution.
- Linux & macOS: Best-effort basis. Startup scripts and hardware extras are maintained, but formal validation is not guaranteed.
- Docker & Google Colab: Explicitly out of scope. Containerized and cloud-hosted environments are not officially supported; users requiring these setups may adapt the source independently (though doing so should be significantly easier and cleaner than before thanks to
uvand deterministic dependency locks).
- Multi-Edition Hardware Adapters:
Providing tailored extras (
cuda,dml,mac,cpu) so NVIDIA, AMD, Intel, and Apple Silicon users each get optimal hardware acceleration. - Streamlined Pure-ONNX Pitch Estimation (RMVPE & FCPE Only): Consolidating pitch estimation exclusively to the gold standard (RMVPE) and a newer, ultra-fast alternative (FCPE) running torch-free via ONNX Runtime and NumPy. See OKF Documentation for details.
- Standalone Portable Packaging:
Automated bundling of isolated Python runtimes and Web UI assets into portable zip packages (
scripts/build_package.py).
For source code development, running automated test suites, code formatting/linting, and building standalone packages, please refer to the dedicated Developer & Build Guide (README_dev.md).
The bundled initial voice models are provided courtesy of their respective original creators. The official Japanese terms of service take precedence; please review the official links before using converted voices:
- Official Website & License: Tsukuyomi-chan Corpus (CV: Rei Yumesaki / © Rei Yumesaki)
- Key Prohibitions:
- Harassing, attacking, or criticizing others.
- Promoting specific political or religious causes.
- Distributing explicit or sensitive content without proper zoning.
- Secondary redistribution of converted voice as raw audio material.
- Official Website & License: Amitaro's Voice Material Workshop
- Key Guidelines:
- You must clearly state that the voice has been converted using Amitaro (or Koharune Ami) and that she is not speaking directly.
- Do not use for defamatory, sensitive, or offensive content violating Amitaro's guidelines.
- Official Website & License: Replica Doll (れぷりかどーる)
- Key Guidelines:
- Please adhere to the character guidelines published on the Replica Doll official site.
- Original Project:
- Original VCClient developed by w-okada.
- Character Voice Corpora:
- Tsukuyomi-chan Corpus (© Rei Yumesaki)
- Amitaro's Voice Material Workshop
- Replica Doll (れぷりかどーる)
The author and contributors assume no liability for any direct, indirect, incidental, consequential, or special damages arising out of the use of or inability to use this software.