Kizuna AI Lab's Sokuji Translates Live Calls Offline Without a GPU
Sokuji is an open-source live speech translator that runs entirely on-device with WASM and WebGPU, no API key or GPU required.
Takeaways
- Sokuji is an open-source live speech translator with 1,000+ GitHub stars, available as desktop app and browser extension
- Local Inference runs ASR, translation, and TTS fully on-device via WASM and WebGPU with no GPU needed
- Bundles 50 ASR models, 55+ translation pairs, and 136 TTS voices across 99+ languages
- Cloud providers include OpenAI Realtime, Gemini, Palabra.ai, Doubao AST 2.0, and OpenAI-compatible endpoints
- Virtual microphone routes translated audio into Zoom, Meet, Teams, Discord, OBS, and other apps
- Licensed AGPL-3.0, built with Electron, React, sherpa-onnx, and Transformers.js
Real-time speech translation has mostly been the domain of cloud APIs with per-minute pricing and privacy tradeoffs. Sokuji, an open-source project from Kizuna AI Lab, takes a different route: it bundles cloud providers and a fully offline on-device pipeline into the same app, so you can translate a Zoom call through OpenAI's realtime API or through a stack of Whisper, Opus-MT, and Piper running in your browser tab.
The repo has crossed 1,000 stars and ships as both a desktop app and a browser extension under AGPL-3.0. The Chrome extension and Edge add-on cover web meetings, while the Electron build targets Windows, macOS, and Linux for any app with mic input.
Offline inference without a GPU
The headline feature is Local Inference. Powered by WASM and WebGPU, Sokuji runs the full ASR to translation to TTS pipeline on-device using your existing CPU and integrated graphics, with no API keys, no internet, and no discrete GPU required. The model catalog is unusually broad for a client-side app:
- 50 ASR models covering 99+ languages, including 32 offline models, 10 streaming models, and 8 WebGPU builds such as Whisper, Cohere Transcribe, and Voxtral Mini 4B