VoxBento

What I built

• Architected VoxBento from scratch as a Google Summer of Code 2026 contributor with FOSSASIA: browser-based, low-latency multilingual interpretation for virtual and hybrid events.
• Interpreters publish over WebRTC/WHIP to MediaMTX and listeners play over WHEP (with HLS fallback); the live audio path never passes through Python.
• FastAPI coordinates events, booths and roles, and runs per-booth speech-to-text workers (faster-whisper, Deepgram and others) fed by RTSP/FFmpeg, with lazy translation and on-demand TTS pushed over WebSockets.
• Transient booth state lives in an in-memory registry with per-booth asyncio locks; durable records persist in PostgreSQL via SQLAlchemy and Alembic. Interpreter handoff happens without listeners changing endpoints.
• Presented the architecture at FOSSASIA Summit 2026 and gave a live demo at Wikimania Paris 2026. The full write-up lives at angarkar.tech/gsoc.

Stack