kizuna-ai-lab/sokuji

What it solves

Sokuji provides real-time, two-way speech translation for bilingual meetings, removing language barriers in live conversations. It allows users to translate their own voice for others to hear and translate incoming audio from other participants into subtitles or speech, regardless of the platform being used (e.g., Zoom, Teams, Discord).

How it works

Sokuji captures audio from the microphone and system audio, then processes it through a pipeline of Automatic Speech Recognition (ASR), translation, and Text-to-Speech (TTS). Users can choose between two processing modes:

• Cloud Mode: Connects directly to providers like OpenAI, Google Gemini, and Soniox for high-performance translation.
• Local Inference: Runs everything on-device using WASM and WebGPU (via sherpa-onnx and Transformers.js), requiring no API keys, internet connection, or dedicated GPU.

Who it’s for

It is designed for people participating in bilingual meetings, gamers, or anyone needing live translation across different applications. It is available as a desktop app for all major operating systems and as a browser extension for web-based meeting platforms.

Highlights

• Two-Way Translation: Auto-detects speaker languages and translates in both directions simultaneously.
• Privacy-First Local AI: Fully offline on-device processing for ASR, translation, and TTS.
• Broad Compatibility: Works with any app via a virtual microphone (desktop) or directly in the browser (extension).
• Extensive Model Support: Access to 50+ ASR models, 55+ translation pairs, and 136 TTS voices.
• AI Audio Tools: Includes built-in noise suppression and echo cancellation.