How svc-develop-team/so-vits-svc works
SoftVC VITS Singing Voice Conversion converts vocal audio while preserving its content and intonation. The principal path is an audio request through the inference entry point, speech-feature and pitch extraction, the conversion model, and waveform decoding; optional retrieval, enhancement, and shallow diffusion extend inference. The repository also includes dataset preparation and model training, plus web and API entry points. The README documents several capabilities whose wiring is not shown in the excerpts, so those links are omitted rather than inferred.
Main entry points: infer_tool.py SVC Inference · models.py SVC Generator · inference_main.py Inference CLI · train.py Model Training
Components 19
User Interfaces
- Inference CLI
inference_main.py - Web UI
webUI.py - Flask API
flask_api.py
Conversion Inference
- SVC Inference
inference/infer_tool.py - Audio Slicing
inference/slicer.py - Speech Encoders
vencoder - Pitch Prediction
modules/F0Predictor - Feature Retrieval
cluster - Shallow Diffusion
diffusion/unit2mel.py - Audio Enhancement
modules/enhancer.py
Conversion Models
- SVC Generator
models.py - Neural Layers
modules/modules.py - Attention Blocks
modules/attentions.py - Waveform Decoders
vdecoder
Training Data
- Feature Preparation
preprocess_hubert_f0.py - Data Utilities
data_utils.py - Model Training
train.py - Diffusion Training
train_diff.py
Other components
- Audio User
Connections 18
- Audio User → Inference CLI: provides audio
- Audio User → Web UI: submits audio
- Audio User → Flask API: submits audio
- Inference CLI → SVC Inference: invokes
inference_main.py - SVC Inference → Audio Slicing: uses
inference/infer_tool.py - SVC Inference → SVC Generator: runs conversion
inference/infer_tool.py - SVC Inference → Speech Encoders: loads encoder
- SVC Inference → Pitch Prediction: selects predictor
- SVC Inference → Feature Retrieval: uses retrieval
inference/infer_tool.py - SVC Inference → Shallow Diffusion: uses diffusion
inference/infer_tool.py - SVC Inference → Audio Enhancement: uses enhancer
inference/infer_tool.py - SVC Generator → Neural Layers: uses layers
models.py - SVC Generator → Attention Blocks: uses attention
models.py - SVC Generator → Waveform Decoders: decodes waveform
models.py - Feature Preparation → Speech Encoders: extracts features
- Model Training → SVC Generator: trains generator
- Data Utilities → Model Training: supplies data
- Diffusion Training → Shallow Diffusion: trains model