VCClient: Real-Time AI Voice Conversion Without a Cloud Dependency

What VCClient Does and Who It Is For

VCClient is a software package that performs real-time voice conversion using AI models. It takes audio from a microphone and converts the voice to match a target speaker, with latency low enough to use in live applications such as streaming, voice chat, or online games. All processing happens on the local machine; no audio is transmitted to a remote server during operation.

The primary users are virtual content creators and streamers who present using character voices, developers who want to build on a voice conversion engine through an HTTP API, and engineers who want to experiment with different AI voice models. The software also runs on Google Colab for users who want to test without a dedicated local GPU.

The repository includes Jupyter notebooks specifically for Colab and Kaggle, which are separate from the standalone application. Linux users get the full capability but must clone and build from source rather than downloading a pre-built binary.

AI Model Backends in v2 and What v1 Supported

VCClient v2 supports two AI backends: RVC (Retrieval-based Voice Conversion) and Beatrice v2. Earlier, version 1 of the application additionally supported MMVC, so-vits-svc, DDSP-SVC, and Beatrice v1, but none of those backends are present in v2.

RVC is a widely-used open voice conversion model that requires a speaker model file trained on the target voice. Beatrice v2 is a separately developed model with its own training code hosted on Hugging Face at the `fierce-cats/beatrice-trainer` repository. The two models differ in how they handle speaker identity and pitch, but the README does not provide a direct technical comparison between them.

As of v2.0.78, version 1 and version 2 of VCClient can run at the same time on the same machine. This lets teams that have existing models trained for the older backends continue using them while testing v2 capabilities. Users who rely on so-vits-svc or MMVC models have no migration path to v2 and must stay on v1 for those models.

Edition Choices and the Platform Matrix

Starting with v2.2.1, VCClient distributes in editions that determine which AI models are available on each platform:

- std edition: Windows, Mac (M1), Linux (x86-64 and aarch64). Supports Beatrice v2 only. - cuda edition: Windows only. Supports Beatrice v2 and RVC. Requires an Nvidia GPU. - onnx edition: Windows and Mac (M1). Supports Beatrice v2 and RVC.

The std edition targets users who only need Beatrice v2 or who do not have an Nvidia GPU. The onnx edition covers Beatrice v2 and RVC on Mac (M1) and on Windows without requiring CUDA drivers. The cuda edition is the only option for Windows users who want RVC with GPU acceleration.

Pre-built binaries for Windows and Mac (M1) are distributed through a Hugging Face repository, not through a package manager or an official GitHub release. A lighter build specifically for Beatrice v2 is available separately at `wok000/light_vcclient_beatrice`. There are no GitHub releases in the repository; versions are tracked through beta and alpha labels in the changelog.

Getting VCClient onto a Linux Machine

Linux users must clone the source repository and follow the build steps in the development documentation. The README notes this directly: Linux is not distributed as a pre-built binary. The clone command is:

git clone https://github.com/w-okada/voice-changer

For anyone who wants to build and run VCClient as a Docker container, the repository includes separate Dockerfiles for the voice changer client, the trainer, and an ONNX converter. The build scripts are defined in package.json. To build the VCClient Docker image:

npm run build:docker:vcclient

This runs the Docker build command targeting `docker_vcclient/Dockerfile` and tags the resulting image as `vcclient`. An equivalent script for the trainer is available as `npm run build:docker:trainer`.

On Mac (M1), the pre-built application is unsigned. The README explains that macOS will display a security warning; users can bypass it by holding the Control key while clicking the application icon. This is a result of Apple's notarization policy, and the README explicitly notes that running it is at the user's own risk.

After installation, users must separately download AI model weight files and point the application to them. The weight files for Beatrice v2 are hosted on Hugging Face at `wok000/vcclient_model`; the README does not include inline instructions for downloading them beyond pointing to the repository.

Standalone Configuration vs. Network Server Mode

VCClient supports two deployment modes. In standalone mode, the conversion engine and the audio client both run on the same PC. This is the typical setup for a single-machine workstation.

The network mode lets the conversion engine run on a separate machine and receive audio over the local network. The machine running the engine does the heavy GPU computation, while the machine running games or other high-load applications sends and receives audio over HTTP. This setup addresses a real constraint: running voice conversion simultaneously with a GPU-intensive game on the same machine can cause frame drops or audio dropouts, because both compete for GPU resources.

The software provides a REST API that other programs can call to adjust conversion parameters without using the web UI. The README notes that curl and any language with an HTTP client library can interact with the API, making it possible to integrate voice conversion control into scripts or other applications.

A known bug in server mode that affected device selection (microphone and speaker) was fixed in v2.0.76. Buffer visualization was added in v2.1.3-alpha to help diagnose latency in the audio pipeline; that feature is currently limited to RVC models.

License Structure and Voice Character Usage Terms

The main repository carries a NOASSERTION license identifier, meaning it does not declare a standard open-source license for the codebase. The AI backends have their own licenses: RVC is covered by its own project's terms, and Beatrice v2 has a custom license documented on Hugging Face.

Three bundled voice character models carry specific usage restrictions. The Tsukuyomi-chan model prohibits using the converted audio to criticize or attack individuals, to promote or oppose specific political, religious, or ideological positions, to publish explicit content without age-gating, or to allow others to use it as raw audio material for further redistribution. Distributing watchable or listenable finished works is explicitly permitted.

The Amitaro model permits voice conversion but requires that any content using the converted voice clearly states that it is Amitaro's voice and not Amitaro speaking. Sensitive or inappropriate content is prohibited under that character's terms.

The Kikyohiroto Mahiro model follows the terms of the Repurikado-ru project. Anyone distributing content that uses these character voices in a commercial or public context should verify the specific character's current terms directly, since the restrictions are detailed and the README references external URLs for the full versions.

Limitations and the Absent Training Pipeline

VCClient does not include a training pipeline for new voice models. To create a custom Beatrice v2 voice, users need the separate Beatrice trainer repository and Colab notebook. To create a custom RVC model, they need the RVC project's trainer. The two training tools are external to VCClient and have their own installation requirements.

The beta and alpha versioning scheme makes it difficult to identify a stable release. Binaries are not on a package registry, so pinning a version for a reproducible deployment requires manually tracking the Hugging Face repository. Not every version is available for every platform: the release notes for v2.0.77 state that it targeted RTX 5090 hardware only, and the developer noted the build was unverified due to not having that GPU.

On Mac hardware that is not an M1 chip (Intel-based Macs), neither CUDA nor ONNX acceleration is available in the pre-built binaries. The README does not document whether the software runs on Intel Macs through the Linux path or through any other method.

The closest alternative with overlapping functionality is the RVC-WebUI project, which also runs RVC models and includes a built-in training interface. The key difference is scope: RVC-WebUI focuses on a single model family, while VCClient supports multiple backends and adds the client-server architecture and REST API. For users who only need RVC and want training integrated into the same tool, RVC-WebUI avoids managing a separate trainer repository.

Editorial conclusion

VCClient suits content creators, streamers, and developers who want fully local AI voice conversion without routing audio through a cloud service. Anyone who needs a trained RVC model should also look at the RVC project's separate trainer, since VCClient does not include a training pipeline. On Windows with an Nvidia GPU, confirm the cuda edition is available for your target version before downloading; the std edition supports only Beatrice v2. The Hugging Face distribution model and the ongoing beta versioning mean there is no stable release on a conventional package registry, which matters if you are building a deployment that needs a pinned version.

Frequently asked questions

Official sources

1. Issues
2. README
3. w-okada/voice-changer on GitHub