How rvc-project/retrieval-based-voice-conversion-webui works
RVC is a voice-conversion application with a WebUI for training and inference, plus a low-latency realtime GUI/VST path. Its principal workflow is selecting a voice model and audio, extracting content and pitch features, optionally retrieving training-set features, then synthesizing converted audio. The repository also includes optional pymss/MSST source separation and a C++ VST worker. Edges rely on README statements and the supplied source index/excerpts; some detailed training and unsampled subsystem wiring is therefore omitted or marked unverified.
Main entry points: webui.py WebUI · pipeline.py Conversion pipeline · rtrvc.py RVC synthesis · train.py Model trainer
Components 24
User interfaces
- WebUI
webui.py - Realtime GUI
realtime_gui.py - Runtime config
configs/config.py
Voice conversion
- VC interface
infer/vc/modules.py - Conversion pipeline
infer/vc/pipeline.py - HuBERT features
infer/hubert.py - Pitch estimators
infer/rmvpe.py - RVC synthesis
infer/rtrvc.py - Synthesizer networks
infer/module/models.py - CUDA graph runtime
tools/cuda_graph.py
Model training
- Training data prep
train/preprocess.py - Feature extraction
train/dataset/extract_hubert_feature.py - Pitch extraction
train/dataset/extract_f0.py - Model trainer
train/train.py - Checkpoint tools
train/process_ckpt.py
Vocal separation
- pymss UI
tools/pymss_webui.py - Stem separation
tools/pymss/separator.py - Separation models
tools/pymss/model_registry.py - Model downloads
tools/pymss/model_download.py
Realtime plugin
- Worker client
RVCRealtimeVST/src/WorkerClient.cpp - Python worker
RVCRealtimeVST/worker/rvc_worker.py
Other components
- User
- Audio host
- Model files
Connections 29
- User → WebUI: operates
README.md - WebUI → Runtime config: loads settings
webui.py - WebUI → VC interface: invokes conversion
webui.py - WebUI → Model trainer: starts training
- WebUI → Checkpoint tools: processes checkpoints
webui.py - WebUI → pymss UI: offers separation
webui.py - User → Realtime GUI: operates
README.md - Training data prep → Feature extraction: prepares audio
- Training data prep → Pitch extraction: prepares audio
- Feature extraction → Model trainer: supplies features
- Pitch extraction → Model trainer: supplies pitch
- Model trainer → Checkpoint tools: saves model
train/train.py - VC interface → Conversion pipeline: runs pipeline
infer/vc/modules.py - Conversion pipeline → HuBERT features: extracts features
infer/vc/pipeline.py - Conversion pipeline → Pitch estimators: estimates pitch
infer/vc/pipeline.py - VC interface → RVC synthesis: loads conversion model
infer/vc/modules.py - RVC synthesis → Synthesizer networks: builds synthesizer
infer/rtrvc.py - Conversion pipeline → CUDA graph runtime: uses acceleration
infer/vc/pipeline.py - RVC synthesis → Pitch estimators: uses pitch estimation
infer/rtrvc.py - pymss UI → Stem separation: runs separation
- Stem separation → Separation models: resolves models
tools/pymss/separator.py - Separation models → Model downloads: requests downloads
tools/pymss/model_registry.py - Model downloads → Model files: stores weights
- Audio host → Worker client: streams audio
- Worker client → Python worker: dispatches blocks
- Python worker → Runtime config: loads settings
RVCRealtimeVST/worker/rvc_worker.py - Python worker → RVC synthesis: runs conversion
RVCRealtimeVST/worker/rvc_worker.py - Worker client → Audio host: returns audio
- WebUI → Model files: uses voice models
README.md