Open-source project
voicepaw/so-vits-svc-fork avatar
voicepaw/so-vits-svc-fork

so-vits-svc-fork: the torch index it pins, the v1.3.2 batch file, and the models it refuses

so-vits-svc fork with realtime support, improved interface and more features.

9,324 stars1,210 forksPythonNOASSERTION

At a glance

What is it?
SoftVC, the so-vits-svc fork with realtime conversion and a PySimpleGUI window, is declared no longer maintained by its own author while its package version sits at 4.2.30. Its install instructions pin a CUDA 12.1 wheel index, its one-click installer is a batch file downloaded from a v1.3.2 tag, and its model compatibility rules contradict each other in consecutive sentences.
Who is it for?
This fork still installs from PyPI with pip, and it remains the shortest path to so-vits-svc style conversion if you already have a 4.0v1 model or a model trained for the fork. It is not a place to plan a long-lived project, because the author has declared it unmaintained, the reasons for that decision are still unaddressed in the dependency list, and the newest published release is v4.2.30 from 2026-02-02.
Can I use it commercially?
Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
Is it still maintained?
Yes. The repository received new commits within the last day.
What is it written in?
Mainly Python, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • voicepaw/so-vits-svc-fork9,324 ★ · -11
Star history
Repository2026-09-132026-11-15
voicepaw/so-vits-svc-fork9.3k9.3k

-11 Stars in the last 63 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

Both manual install routes pin the same cu121 wheel index

The plain pip route is three commands, and the middle one exists on purpose:

shell
python -m pip install -U pip setuptools wheel
pip install -U torch torchaudio --index-url https://download.pytorch.org/whl/cu121 # https://download.pytorch.org/whl/nightly/cu121
pip install -U so-vits-svc-fork

That second line pins torch and torchaudio to a CUDA 12.1 wheel index and keeps the nightly equivalent in a trailing comment on the same line. If you have no GPU, or you are on macOS, that line is the one to delete, and the note beside it says MPS is probably supported. On Linux with an AMD card you substitute https://download.pytorch.org/whl/nightly/rocm5.7 for the index, and AMD on Windows is not supported at all, tracked as issue 120.

The pipx route hits the same pin with one extra step, because pipx hands the package its own isolated environment:

shell
pipx install so-vits-svc-fork --python=3.11
pipx inject so-vits-svc-fork torch torchaudio --pip-args="--upgrade" --index-url=https://download.pytorch.org/whl/cu121 # https://download.pytorch.org/whl/nightly/cu121

Both routes are labelled manual and the pipx one is labelled experimental. Whichever you pick, the torch index is your decision to make, not the package's.

The one-click installer is downloaded from a v1.3.2 tag

Option 1 is a link to a Windows batch file, and the link itself carries the whole specification: it resolves to the releases download path for tag v1.3.2 and to a file named install.bat. The text says only that this BAT file will automatically perform the steps described below.

Compare that against the rest of the repository. pyproject.toml declares version 4.2.30, the newest release tag is v4.2.30 published on 2026-02-02, and the source for the script sits in the easy-installation/ directory rather than at the root. So the one-click path is bound to a script published under a tag from the older 1.x numbering, and nothing at the download URL states which package version it installs. The README also gives the pipx route a development-only pipx install on Windows, citing pypa/pipx issue 940, which tells you the maintainers hit friction getting this path to work.

The consequence for a reader is narrow but real: if you take the batch file, you are running someone else's idea of the install sequence from an old tag, on Windows only, and you cannot tell from the link what ends up on disk. On Linux and macOS the venv and conda blocks are the documented path, and both of those pin Python 3.11.

The summary says 4.1 models are both supported and not supported

The opening description of the fork contradicts itself within two sentences. It says the project is based on branch `4.0` (v1) (or `4.1`) and the models are compatible, then says `4.1` models are not supported. The Notes block under Usage restates the rule in different words: models other than for 4.0v1 or this repository are not supported.

Read together, the usable compatibility set is 4.0v1 models plus models trained for this fork, and 4.1 is named in the sentence that includes it and in the sentence that excludes it. That matters because a model file is the one thing you cannot generate yourself here. The feature list says pretrained models are downloaded automatically and that you no longer need to install fairseq, and it points to Hugging Face and to the so-vits-svc-fork tag on CIVITAI as the places to get them.

A model whose page does not name the base version it was trained on is hard to check before you build an inference pipeline around it.

The reasons for stopping name Typer and PySimpleGUI, both still in the manifest

The No Longer Maintained section gives four reasons. Two of them are about the interface: PySimpleGUI is no longer LGPL, and using Typer is getting more popular than directly using Click. The other two are about the field moving on, with many better alternatives now available, and about the author not having the skills, time or money to build the more modular and easier-to-install repository that was the goal.

The dependency list has not caught up with those reasons. pyproject.toml still requires click>=8.1.8 and pysimplegui-4-foss>=4.60.4.1, and no typer appears in the requirements at all. The console entry points confirm which stack you actually get: `scripts.svc` points at so_vits_svc_fork.__main__:cli and `scripts.svcg` points at so_vits_svc_fork.gui:main.

So the migration the author describes as a reason to stop never happened in the manifest. The Click command line and the PySimpleGUI window are still the interface, which means the two stated complaints are still conditions you inherit, and a contributor reading the reason list would have no idea the code never moved.

The Dockerfile pins torch 2.0.1 and starts the GUI, not the CLI

The image is six instructions on a digest-pinned base:

dockerfile
FROM pytorch/pytorch:2.0.1-cuda11.7-cudnn8-runtime@sha256:82e0d379a5dedd6303c89eda57bcc434c40be11f249ddfadfd5673b84351e806
RUN ["apt", "update"]
RUN ["apt", "install", "-y", "build-essential"]
RUN ["pip", "install", "-U", "pip", "setuptools", "wheel"]
RUN ["pip", "install", "-U", "so-vits-svc-fork"]
ENTRYPOINT [ "svcg" ]

Three things do not line up with the rest of the project. The base image carries pytorch 2.0.1 on CUDA 11.7, while pyproject.toml requires torch>=2.8.0 and torchaudio>=2.8.0, and the pip instructions elsewhere use the cu121 index, so the install line resolves a newer torch on top of the base image and on a different CUDA generation than the pinned base. The entrypoint is `svcg`, which launches the GUI, while the documented command line entry points are:

shell
svcg
shell
svc vc
shell
svc infer source.wav

`svc vc` gives realtime conversion from a microphone and `svc infer` handles a file, and neither is what the container starts. Combined with the WSL note that the GUI will not work without finding an audio device, this makes the image unusable as a headless server unless you override the entrypoint and solve audio yourself.

MIT in the manifest, undetermined on the repository, pre-alpha in the classifier

Three pieces of metadata disagree with each other. The repository license is reported as undetermined, while pyproject.toml declares `license = { text = "MIT" }` and a LICENSE file sits at the root, so the licence text is in the repository even though the repository page does not identify it. The classifier set says Development Status :: 2 - Pre-Alpha on a package whose version is 4.2.30, and the feature list claims roughly twice faster training than the original project, so the version number tells you nothing about stability.

The python story is looser still. `requires-python` is >=3.9 and the classifiers enumerate 3.9 through 3.13, while the pipx line hard-codes `--python=3.11` and both the venv and the conda blocks create Python 3.11 environments. Anyone on 3.9 satisfies the manifest and has no documented install path.

For redistribution this matters more than for casual use: if you are shipping the package inside a product, read the LICENSE file rather than the repository page, and pin the interpreter to 3.11 because that is the version every documented route was written against.

Pip upgrade stops at the February release while commits keep arriving

Updating is one active line and one commented line:

shell
pip install -U so-vits-svc-fork
# pipx upgrade so-vits-svc-fork

The release history is v4.2.30 on 2026-02-02, v4.2.29 on 2025-10-27 and v4.2.28 on 2025-10-26, while the last push to main is dated 2026-11-02. Commits are landing on the branch and no newer release has been cut since February, so the upgrade command moves you to v4.2.30 and stops. Everything after that exists only for anyone building from source.

The README also names its own replacements with their licences and their stated states: IAHispano/Applio under MIT, marked actively maintained; fumiama's RVC under AGPL; the original RVC under MIT, marked no longer maintained; w-okada's VCClient, which offers a web-based GUI for realtime conversion but is described as not quite actively maintained; fish-diffusion, which tried to be quite modular but is not actively maintained; yxlllc/DDSP-SVC with releases issued occasionally; yxlllc/ReFlow-VAE-SVC; and coqui-ai/TTS, which was for TTS rather than conversion and is no longer maintained. The author also advises taking social networking posts about new voice changers with semi-doubt. If you want maintenance rather than this fork's feature set, the README points at Applio.

Editorial conclusion

This fork still installs from PyPI with pip, and it remains the shortest path to so-vits-svc style conversion if you already have a 4.0v1 model or a model trained for the fork. It is not a place to plan a long-lived project, because the author has declared it unmaintained, the reasons for that decision are still unaddressed in the dependency list, and the newest published release is v4.2.30 from 2026-02-02. Before you commit, check your model's base version against the one supported model family, decide which torch index your GPU actually needs, and read the Dockerfile, which pins a different torch and CUDA version than the pip instructions.

Frequently asked questions

Can I clone my voice locally with so-vits-svc-fork?

This package performs voice conversion rather than voice cloning from a text prompt. Inference runs either from a file with `svc infer source.wav` or in realtime from a microphone with `svc vc`, and the GUI is launched with `svcg`. Pretrained models are available on Hugging Face or under the so-vits-svc-fork tag on CIVITAI, and inference runs on your own machine.

How do I install so-vits-svc-fork on Linux or macOS?

Create a Python 3.11 environment with `python3.11 -m venv venv` and `source venv/bin/activate`, or `conda create -n so-vits-svc-fork python=3.11 pip` followed by `conda activate so-vits-svc-fork`, then run `python -m pip install -U pip setuptools wheel`, the torch line pinned to https://download.pytorch.org/whl/cu121, and `pip install -U so-vits-svc-fork`. If you have no GPU or are on macOS, delete the torch line.

What models does so-vits-svc-fork support?

Models other than for 4.0v1 or trained for this repository are not supported. The project states that it is based on branch 4.0 (v1) (or 4.1) with compatible models, then immediately says 4.1 models are not supported, so the practical set is 4.0v1 models plus models made for the fork.

Can so-vits-svc-fork run on an AMD GPU?

On Linux, yes, by swapping the wheel index: the note says to replace `--index-url https://download.pytorch.org/whl/cu121` with `--index-url https://download.pytorch.org/whl/nightly/rocm5.7`. AMD GPUs are not supported on Windows, tracked as issue 120, and GPU inference requires at least 4 GB of VRAM.

Is so-vits-svc-fork still maintained?

The README carries a No Longer Maintained section stating that updates have been limited to maintenance since Spring 2023. The last release tag is v4.2.30 from 2026-02-02 even though the last push to main is dated 2026-11-02, and the README points at IAHispano/Applio, marked actively maintained, among its alternatives.

Official sources

  1. Issues
  2. README
  3. Releases
  4. voicepaw/so-vits-svc-fork on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/voicepaw-so-vits-svc-fork.svg)](https://hysenlabs.com/projects/voicepaw-so-vits-svc-fork)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/voicepaw-so-vits-svc-fork.svg)](https://hysenlabs.com/projects/voicepaw-so-vits-svc-fork)

Was this useful?