2.3.260718
Chainlit/chainlit2.3.260718Jul 21, 2026by RVC-Boss
AI Summary
Major update to the Retrieval-based Voice Conversion WebUI featuring new algorithms, backend changes, and performance optimizations including CUDA Graph acceleration.
Key Highlights
- Supports FCPE pitch extraction algorithm
- Real-time VST plugin supports CUDA Graph acceleration (up to 4.7x speedup)
- Backend for vocal separation changed from UVR5 to PyMSS for better results
- Automatic GPU detection for inference mode and precision
- Fixes dependency issues (pyworld, fairseq, matplotlib, pyav, onnxruntime-gpu)
New Features
- FCPE pitch extraction algorithm support
- CUDA Graph acceleration for real-time VST mode
- PyMSS backend for vocal separation
- GPU auto-detection for inference mode
- Single-card training without DDP
Full Release Notes
### 完整包 Complete package For Nvidia GPU users (<50x0): https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/RVC20260718Nvidia.7z For Nvidia GPU users (>=50x0): https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/RVC20260718Nvidia50x0.7z For AMD/Intel GPU users: https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/RVC20260718AMD_Intel.7z RVC Real-Time VST2/VST3 Plugin (For user instructions, please refer to the README included in the archive. You should unzip the complete package first.) https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/RVCRealtime-Win64VST.zip 中国用户可使用以下2个源加速下载 1、免费满速下载直链合集:https://www.yuque.com/flowercry/hxf0ds 2、度盘需要超级会员才可不限速下载 链接:https://pan.baidu.com/s/16BHeTXRq2aGUzDZo1BY1Lw 提取码: tvy6 Changelog [(English verison)](https://github.com/liujing04/Retrieval-based-Voice-Conversion-WebUI/blob/main/docs/en/Changelog_EN.md): 降低网页使用难度(推理输入音频无需输入路径,前端实时日志优化,保存推理用小模型后自动刷新模型列表,推理首页选择模型后自动检测匹配的index路径等) 数据特征提取、训练模型、索引阶段可中途掐断,重启继续进度 GPU自动识别判断推理模式和精度 修复依赖问题(例如pyworld,fairseq,matplotlib,pyav,onnxruntime-gpu环境修复等) 解决编码问题 网页启动自动探寻可用端口 单卡训练不用DDP 基础模型不变,若干影响效果的修复,使用现有微调模型更新整合包/代码包版本重新推理,或者使用新版本重新训练再推理,均有可能效果会有提升。 支持FCPE音高提取算法 实时变声模式gui支持cuda graph推理加速,算法推理延时指标最高加速4.7倍(4090D测试结果),同时显卡占用相对降低20~30%左右 移除了已明显无用或过时的功能 人声与伴奏分离后端由 UVR5 更换为 PyMSS,并支持推理时中断 (更好的人声分离算法能显著提高推理效果) 优化输入音频的读取与重采样流程,在条件允许时使用 GPU 处理,以提高推理效率并降低 CPU 占用 修复 Windows 环境下训练进程的 libuv 兼容问题