ShibaSVC

ShibaSVC is a local singing-voice conversion toolkit focused on practical HiFiSinger training and inference. It provides a Gradio GUI and command-line tools for dataset preparation, compact feature extraction, training, and conversion.

Highlights

• HiFiSinger with HiFiGAN, RefineGAN, the original Wavehax, and Shiba HarmonicRaw decoder configurations.
• ContentVec content features, plus RMVPE and other pitch extractors.
• Local-first model assets; opening the GUI never starts a download.
• Physical batch 8 profiles designed for an 8 GB GPU, including branchwise HarmonicRaw discriminator training.
• Tensor-only Safetensors exports with embedded resolved configuration.
• Explicit safe checkpoint loading with PyTorch's tensor-oriented weights_only policy.
• Rich terminal progress and structured log/progress events in the GUI.

Quick start

Python 3.12 is required. Installing from source is recommended for GPU users because the provided installer selects the PyTorch wheel index explicitly. On Linux:

./install.sh
./run.sh

On Windows, run install.bat and then run.bat. Native Windows uses eager CUDA for maximum compatibility; use WSL2 when torch.compile fusion is required. For CPU-only PyTorch, use ./install.sh --cpu or install.bat --cpu.

After the first PyPI release, the packaged GUI can instead be installed with python -m pip install shiba-svc and launched with shiba-svc. Install the desired CPU/CUDA build of PyTorch first. See the installation guide for the complete PyPI, GitHub source archive, CUDA, and Windows instructions.

The GUI contains Conversion, Dataset preparation, Feature extraction, and Training tabs. It discovers datasets below assets/datasets/ and checkpoints in the local model directories. Pretrained checkpoints can be placed in pretrained/ or selected by path in the Training tab.

Documentation

• Installation from PyPI or source, GPU setup, and development
• Dataset preparation and feature extraction
• Training, checkpoints, and decoder profiles
• Troubleshooting
• Preflight, manifests, Safetensors, and release publishing
• Command-line reference

Inference

Safetensors inference exports contain their model and preprocessing configuration:

shiba-svc-infer \
  --checkpoint logs/my_voice/epoch=59-step=13520-valid_loss=0.92.inference.safetensors \
  --input input.wav \
  --output output.wav

Pass --config for a legacy or full checkpoint without embedded metadata. Inference supports pitch adjustment, ContentVec extraction, independent input/output normalization, reference-style strength, and optional target-content retrieval. Input normalization defaults to the checkpoint's training-compatible setting (post_peak for current and legacy checkpoints). CUDA resources are released after each GUI conversion.

Repository layout

shiba_svc/                 # Python package
  configs/                 # packaged model/training configurations
  gui/                     # Gradio application and visual assets
  cli/                     # installed command-line implementations
assets/datasets/           # source datasets used by the GUI
logs/                      # prepared data, TensorBoard, and checkpoints
pretrained/                # explicitly downloaded/cached model assets
examples/                  # manual workflows requiring local assets
docs/                      # task-oriented documentation

Tests

python -m pytest -q

Manual model and visualization workflows that require external checkpoints or audio live under examples/ and are not collected by pytest.

Attribution and responsible use

ShibaSVC is based on the Fish Diffusion codebase. Shiba HarmonicRaw is an independent decoder influenced by the alias-control ideas described in Aliasing-Free Neural Audio Synthesis; it is not the official Pupu-Vocoder implementation. See Third-party notices for the exact relationship and upstream licenses.

Use only audio for which you have the necessary rights and permissions. Voice conversion can be misused for impersonation; disclose generated or converted audio where appropriate and follow applicable laws and platform rules.