VCClient ships models by edition, builds by Hugging Face, and leaves the RTX 5090 unverified
Realtime Voice Changer.
At a glance
- What is it?
- w-okada/voice-changer is VCClient, a realtime AI voice changer written mostly in Python. Its behaviour depends on which edition you install, its binaries live on a model host rather than on a release page, its Linux path is a clone, and one GPU build was published by a developer who does not own that GPU.
- Who is it for?
- VCClient fits you if you have already decided which AI voice model you want and which edition carries it, and if you are willing to install from a Hugging Face repository instead of a release page. It does not fit you if you need one build to cover every model family, or if a signed installer matters to you on macOS.
- 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 last received commits 50 days ago.
- 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.
Star history
- w-okada/voice-changer21,382 ★ · +355
| Repository | 2026-09-21 | 2026-11-15 |
|---|---|---|
| w-okada/voice-changer | 21k | 21.4k |
355 Stars in the last 55 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
There are no GitHub releases, so the binaries come from a model host
The repository publishes no releases at all, which removes the usual way to answer what to download. Instead the Windows and M1 Mac versions are pulled from Hugging Face, from the VCClient repository at wok000/vcclient000 and a second one for the lightweight Beatrice v2 build at wok000/light_vcclient_beatrice. Linux users are told to clone the repository and use it from there. No checksum, no file listing and no version history sits in this tree to check any of it against. What the changelog does tell you is that the set of release editions changes per version: v.2.2.2-beta ships std_win, std_mac and std_lin_aarch64, and it warns from that version onward that supported models differ by edition and that not every edition ships on every version.
The edition table decides your model set, and std is Beatrice only
From v.2.2.1 the editions carry different models, and the table is short enough to memorise. std runs on win, mac(m1), linux(x86-64) and linux(aarch64) and supports Beatrice alone. cuda exists on win only and adds RVC. onnx exists on win and mac(m1) and adds RVC. There is no cuda edition for Mac and no onnx edition for Linux, so an M1 user who wants RVC is on the onnx build whether or not CUDA exists for that machine. Across the two major versions the picture shifts again: MMVC, so-vits-svc and DDSP-SVC are marked n/a under v.2 and supported under v.1, while Beatrice v1 is the reverse, supported only on Windows in v.1. Moving from v.1 to v.2 costs you three model families unless you stay on a v.1 line.
The RTX 5090 build shipped without anyone owning an RTX 5090
The changelog is unusually candid about one release. v.2.0.77-beta is labelled as only for the RTX 5090 and experimental, and it states that the related modules were updated for 5090 support and that the developer does not own an RTX 5090, so operation is unverified. That is the honest version of a compatibility claim, and it means on that card you are the test. The surrounding entries show which paths are young. v.2.0.78-beta avoids an RVC model upload error, allows running side by side with ver.1.x, and adds more selectable chunk sizes. v.2.0.76-beta adds Beatrice speaker merge and auto pitch shift and fixes a device selection fault in server mode. v.2.0.73-beta fixes pitch and formant changes not being reflected in Beatrice v2, plus a failure on ONNX models built with an Applio embedder.
The Mac build is unsigned, and every launch needs a modifier key
There is a section explaining that the software is not signed by its developer. A warning appears, and holding the control key while clicking the icon lets it run. The stated reason is Apple's security policy, and the stated responsibility is yours. The same posture shows in the disclaimer, which declines any liability for direct, indirect, consequential or special damage arising from use or inability to use the software. Between the two, the project's position is that the operating system gate and the user carry the risk.
Linux is a clone, and Google Colab is four notebooks
Supported platforms are listed as Windows, Mac(M1), Linux and Google Colab, but the delivery differs per platform in a way the list does not convey. Windows and M1 Mac are downloads, Linux is a clone of this repository, and the Colab route is not an install at all. Four notebooks sit at the repository root for that last case: Kaggle_RealtimeVoiceChanger.ipynb, Realtime_Voice_Changer_on_Colab.ipynb, and two Hina_Mod variants for Kaggle and Colab. The consequence for you is that your install shape determines your upgrade story. A downloaded build moves when the model host updates it, while a clone moves when you pull, and a notebook moves when someone rewrites the cell.
The npm package is a Docker build wrapper, and one script points at a missing directory
package.json names four image builds and three push scripts, and three of the four image builds stamp a timestamped dummy file before invoking the builder.
"build:docker": "date +%Y%m%d%H%M%S > docker/dummy && DOCKER_BUILDKIT=1 docker build -f docker/Dockerfile docker/ -t voice-changer",
"build:docker:onnx": "DOCKER_BUILDKIT=1 docker build -f docker_onnx/Dockerfile docker/ -t onnx-converter",The root holds docker/, docker_folder/, docker_trainer/ and docker_vcclient/, but no docker_onnx/, so the onnx converter script has no context directory at this level. Two more signals sit alongside: the test script is a stub that reports no test specified and exits 1, so nothing runs automatically, and the version field says 1.0.0 while the application is at v.2.2.2-beta. The declared license is ISC, which sits oddly beside three root license files.
Server mode is the documented interface, and the failure docs are Japanese only
Two deployment shapes are supported: a self contained conversion on your own machine, and conversion over the network. The stated reason for the second is load, since running it across a network offloads the conversion work externally when you are also running something demanding such as a game. The interface is a REST API, which lets you write a client in any language, and it can also be driven with an HTTP client already present in the operating system such as curl. That makes the documented interface a network one rather than a file pipeline, which puts communication between the two halves at the centre of your support needs. The troubleshooting page linked from the readme is a Japanese guide to communication problems, tutorials/trouble_shoot_communication_ja.md.
Model licenses are per row, and the bundled character voices carry usage bans
License terms are not uniform. RVC, MMVC, so-vits-svc and DDSP-SVC all point you at their own repositories, while both Beatrice versions are marked with a proprietary license of their own. So what you are licensed to do depends on which model file you load, not on this project's own terms. On top of that, the character voices bundled for text to speech have their own restrictions. The Tsukuyomi-chan voice changer follows its corpus terms and bans using the converted audio to criticise or attack people, to call for agreement or opposition to a political, religious or ideological position, to publish strongly stimulating expressions without zoning, and to publish it in a form that permits others to reuse it as material. Distribution as a work for viewing is fine. Amitaro's rules additionally require stating that the voice has been converted and that she is not the speaker.
Editorial conclusion
VCClient fits you if you have already decided which AI voice model you want and which edition carries it, and if you are willing to install from a Hugging Face repository instead of a release page. It does not fit you if you need one build to cover every model family, or if a signed installer matters to you on macOS. Before you install, check the edition table for your platform, read the license row for the specific model, and remember that the RVC model set is the only one spanning both the v.1 and v.2 lines.
Frequently asked questions
how to install VCClient, the voice changer?
Windows and M1 Mac builds are downloaded from Hugging Face, either the VCClient repository at wok000/vcclient000 or the light build for Beatrice v2 at wok000/light_vcclient_beatrice. Linux users are told to clone this repository instead, and Google Colab users work from the notebooks at the repository root.
how to use VCClient for voice changing?
It converts audio in realtime with AI models, and it runs either standalone on your machine or across a network so the conversion load can sit on another machine. A REST API is provided so clients can be written in any programming language, and it can be driven with an HTTP client built into the operating system such as curl.
which AI models does VCClient support in each version?
Under v.2, RVC and Beatrice v2 are supported, while MMVC, so-vits-svc and DDSP-SVC are marked n/a. Under v.1, RVC, MMVC, so-vits-svc and DDSP-SVC are supported and Beatrice v1 is supported only on Windows. From v.2.2.1 the available models also depend on the edition, where std carries Beatrice alone and the cuda and onnx editions add RVC.
how do I use VCClient with Discord?
The README gives no Discord specific guidance. What it describes is server mode, where conversion runs over the network and the load can be moved off the machine running a demanding application, driven through the REST API from any HTTP client including curl.
Are voice changers illegal under VCClient's terms?
The project's own terms restrict particular uses rather than declaring the software illegal. The bundled Tsukuyomi-chan voice changer bans using converted audio to criticise or attack people, to call for agreement or opposition to a specific political, religious or ideological position, to publish strongly stimulating expressions without zoning, and to publish it in a form that lets others reuse it as material, and the project disclaims liability for resulting damage.
Official sources
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.
[](https://hysenlabs.com/projects/w-okada-voice-changer)Star history chart for your README
[](https://hysenlabs.com/projects/w-okada-voice-changer)Was this useful?