RVC-Project/Retrieval-based-Voice-Conversion-WebUI

A new command-line interface (CLI) is introduced for offline RVC inference, allowing users to process audio files or directories with options for speaker ID, pitch shift, F0 method, and output format. The \infer\ module is significantly expanded with new files for audio handling (\audio.py\), CLI logic (\cli.py\), and model components (\fcpe.py\, \hubert.py\, \module/...\). Audio loading and resampling now support GPU acceleration via \torchaudio\ when available, falling back to FFmpeg/CPU otherwise. Additionally, CUDA Graph support is added to accelerate FCPE and HuBERT inference, improving performance on compatible GPUs.

infer · high confidence

A new Windows x64 VST2 and VST3 plugin project has been added to the repository. The plugin implements real-time voice conversion using C++17 and iPlug2, with model loading and RVC inference handled by a separate Python worker process. The project includes build scripts, test scripts, and bilingual developer guides (English and Simplified Chinese) to help developers compile and use the plugin.

RVCRealtimeVST · high confidence

The application now includes dedicated launch scripts and Python entry points for both a real-time voice conversion interface (realtime\_gui.py) and an offline WebUI (webui.py). The offline WebUI defaults to disabling CUDA Graphs for stability, while the real-time GUI supports GPU processing for audio loading, resampling, and inference to improve efficiency. The project also introduces a new dependency on the pymss library for vocal separation, with specific requirements files (requirments\_cu118\_py312.txt, requirments\_cu128\_py312.txt, requirments\_cpu\_py312.txt) that configure PyTorch, ONNX Runtime, and other libraries for CUDA 11.8, CUDA 12.8, and CPU/DirectML environments respectively.

(repo-wide) · high confidence

New configuration files have been added for model versions v1 and v2 at sampling rates of 32k, 40k, and 48k, defining training and model parameters. Additionally, the system now supports CUDA Graph inference acceleration, with the configuration module automatically detecting and enabling CUDA Graphs on compatible hardware to improve performance.

configs · high confidence

Introduces a new internationalization system for the training interface, providing English, Spanish, French, and Italian translations for all UI strings related to model training, data preprocessing, and feature extraction.

i18n · high confidence

Added new modules for vocal separation and bandit-based audio splitting. The vocal remover now includes a comprehensive set of model metadata, a common separator base class, and a full MLX backend for MPS devices, alongside standard PyTorch implementations. Additionally, a new bandit module was introduced, featuring spectral components, band splitting logic, and a multi-source multi-mask RNN model for separating audio into different frequency bands.

(repo-wide) · high confidence

The tools directory now includes a new 'pymss' package that provides a Python API for music source separation, including model cataloging, downloading, audio I/O, and ensemble utilities. This replaces the previous UVR5 separation backend. Additionally, a new 'cuda\_graph.py' module adds CUDA Graph inference acceleration support, while 'process\_utils.py' adds robust process tree termination, 'file\_io.py' adds safe text file reading, and 'multispeaker.py' adds manifest building for multi-speaker training sets.

tools · high confidence

Added a complete set of scripts and utilities for multi-speaker voice cloning training. This includes \train.py\ for the main training loop with distributed data parallel support, \data\_utils.py\ for loading and batching audio/text pairs, \preprocess.py\ for audio slicing and resampling, \extract\_f0.py\ and \extract\_hubert\_feature.py\ for pitch and HuBERT feature extraction, \train\_index.py\ for FAISS index training, and supporting modules for loss functions, mel processing, and checkpoint management.

train · high confidence

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

• CAI 45 → 48 (+3.3)
• Rubric changed (rubric-2026.08.15 → rubric-2026.09.15) — scores are not directly comparable.

• Code Health 91 → 84 (-6.5)
• Architecture 94 → 99 (+5.2)
• Maturity 60 → 66 (+6.1)
• Readiness 15 → 17 (+2.6)
• Security 77 → 82 (+5.2)

• Dimension evaluation failed
• Duplicated block (13 lines × 2) (train/dataset/slicer2.py)
• Duplicated block (17 lines × 2) (infer/module/models.py)
• Duplicated block (5 lines × 2) (infer/fcpe.py)
• Duplicated block (5 lines × 2) (infer/fcpe.py)
• Duplicated block (5 lines × 2) (infer/rtrvc.py)
• Duplicated block (7 lines × 2) (tools/pymss_core/modules/bs_roformer/bands.py)
• Duplicated block (9 lines × 2) (tools/pymss_core/modules/bs_roformer/bands.py)
• Low: security finding (details withheld)
• Low: security finding (details withheld)
• Medium: security finding (details withheld)
• Medium: security finding (details withheld)
• No exposed public API
• The README covers Studio One VST2/VST3 integration but does not state the exact minimum RVC runtime version required (e.g. Python 3.8 vs 3.10) or whether it supports newer versions. (RVCRealtimeVST/resources/README.txt)
• early-stage repository — too little history to judge knowledge freshness
• git history depth insufficient
• single-maintainer — knowledge-concentration (bus factor) risk

• Attention._cuda_or_default_attention (cognitive 18) (tools/pymss_core/modules/bs_roformer/transformer.py)
• Documentation: no architecture or design documentation (docs/jp/faiss_tips_ja.md)
• Documentation: no architecture or design documentation (docs/jp/faq_ja.md)
• Documentation: no project overview (docs/jp/README.ja.md)
• Duplicated block (10 lines × 2) (infer/module/models.py)
• Duplicated block (10 lines × 2) (tools/pymss/modules/vocal_remover/vr_mlx.py)
• Duplicated block (10 lines × 2) (tools/pymss/workflow.py)
• Duplicated block (10 lines × 2) (tools/pymss_core/modules/demucs4ht.py)
• Duplicated block (10 lines × 2) (tools/pymss_core/modules/demucs_local.py)
• Duplicated block (10 lines × 2) (tools/pymss_core/modules/demucs_mlx.py)
• Duplicated block (10 lines × 2) (tools/pymss_core/modules/legacy_demucs.py)
• Duplicated block (10 lines × 2) (train/data_utils.py)
• Duplicated block (10 lines × 3) (tools/pymss_core/modules/demucs_mlx.py)
• Duplicated block (10–11 lines × 2) (tools/pymss_core/modules/bs_roformer/bands.py)
• Duplicated block (10–11 lines × 3) (webui.py)
• Duplicated block (11 lines × 2) (tools/pymss/cli.py)
• Duplicated block (11 lines × 2) (tools/pymss_core/modules/demucs_local.py)
• Duplicated block (11 lines × 2) (tools/pymss_core/modules/demucs_local.py)
• Duplicated block (11 lines × 2) (tools/pymss_core/modules/demucs_local.py)
• Duplicated block (11 lines × 2) (tools/pymss_core/modules/demucs_local.py)
• …and 189 more

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

RVC-Project/Retrieval-based-Voice-Conversion-WebUI was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.