Local-first realtime Russian voice avatar with a browser VRM interface.
Speak Russian. Get a local voice response from an animated avatar.
Rula connects local speech recognition, a Qwen language model, local speech synthesis and audio-driven facial animation through LiveKit. The browser renders a 3D VRM avatar; inference runs on a separate Windows / WSL2 / NVIDIA GPU workstation.
- Voice interface: Russian speech recognition, synthesis and interruption handling.
- Local inference: STT, LLM, TTS and face animation run in the local runtime.
- Realtime contracts: generation IDs, bounded playback and audio-face timing.
- Open architecture: FastAPI agent, React / Three.js client and shared event protocol.
Development preview. Rula is a demo / early runtime. GPU, latency, soak, licensing and air-gap qualification remain acceptance gates. The project website is independent of the voice backend. The former public interactive demo at ai.igor-ya.ru is currently unavailable.
flowchart LR
Browser["Browser · voice + VRM"] <-->|LiveKit| Agent["Local voice agent"]
Agent --> STT["Russian STT"]
Agent --> LLM["Qwen LLM"]
Agent --> TTS["Local TTS"]
TTS --> Face["Audio2Face · blendshapes"]
Face --> Browser
git clone https://github.com/ykshv/rula.git
cd rulaFollow the setup guide for the GPU host, model downloads and local services. The default runtime uses local inference; Internet access is needed to obtain dependencies and model artifacts.
| Component | Current evidence |
|---|---|
| Public project website | Static documentation; independent of the interactive demo |
| Local voice-avatar runtime | Demo implementation; verify readiness on the target workstation |
| Production and air-gap use | Pending the acceptance gates |
Created and maintained by Igor Yakushev. Read the architecture article On-prem voice agents (Russian).
Related project: Rokid Physical AI uses smart glasses as a voice-and-camera interface for robot arms.
Documentation · Русская инструкция · Architecture · Contributing · License
MIT covers this repository's code and documentation. Third-party models, avatar assets and external runtimes retain their own licenses. See SECURITY.md for reporting and scope.
