Browser-First Interpreter Booth Keeps Python Off the Audio Path

What it does

Voxbento is a real-time interpretation platform for live events built around a simple idea: interpreters should need nothing but a browser. An interpreter monitors the conference floor via an embedded Jitsi iframe and speaks into their mic; the audio travels via WebRTC/WHIP to MediaMTX and reaches attendees through WHEP playback with sub-second latency. A FastAPI portal handles booth coordination—active speaker handoffs, relay chat, and JWT auth—while background jobs transcribe and translate floor audio using whatever AI provider you have plugged in.

The interesting bit

The architecture deliberately keeps Python out of the hot path: the interpreter’s audio hits MediaMTX directly via WebRTC, so the FastAPI backend never touches an RTP packet. That separation lets the platform bolt on transcription and translation—via Deepgram, OpenAI, Groq, Anthropic, Gemini, or optional NVIDIA Riva—without adding latency to the broadcast stream. There is even a client-side delay knob so organizers can align translated audio with delayed external video feeds.

Key highlights

• Browser-first, zero-install workflow for interpreters; attendees listen via WHEP.
• Audio ingest uses WHIP, playback uses WHEP, both terminating at MediaMTX for sub-second latency.
• FastAPI backend manages booth state, handoffs, and chat over WebSocket, not audio.
• Background transcription and translation support multiple third-party and local AI models.
• Optional listener-side audio delay (up to 8 s) to sync with lagging video or captions.

Caveats

• The README is thin on how the background transcription and translation are exposed to end users; it is unclear whether those are real-time captions, post-event logs, or side channels.
• It is an integration-heavy stack—Jitsi, MediaMTX, Docker, and multiple external AI keys are mandatory moving parts, not optional extras.

Verdict

Worth a look if you are building live event infrastructure and need an open, self-hosted alternative to proprietary interpretation platforms. Skip it if you were hoping for a lightweight drop-in script; this is a full-stack appliance.