Project 01 · Live prototype
Open-Air Real Time Translator
An open-air, two-way voice translator so neighbors who speak any language can simply understand each other — a small, modern echo of Pentecost.
“Each one heard them speaking in his own language.”
At Pentecost, a crowd from every nation each heard the good news in their own native tongue. This project is a small, modern echo of that moment: an open-air, two-way voice translator.
Two people who don't share a language stand near one device. One speaks; it speaks the translation out loud; the other replies; it translates back. No headphones, no app for the guest, no walkie-talkie button — just a conversation. In our multicultural South Florida, where English, Spanish, and Haitian Creole live side by side, that's a quiet way to love your neighbor.
What we're trying to do
Where it helps
At the counter
A Deerfield Beach shop owner gets a customer who speaks only Spanish or Haitian Creole. Now they can serve them warmly in their own language — no app for the customer to install, no awkward hand signals.
In the congregation
Haitian Creole–speaking members hear the sermon in their own language as it's preached — without pulling a volunteer interpreter out of worship every week.
Across the neighborhood
Care visits, school meetings, the welcome table — anywhere two neighbors want to understand each other but don't share a language, a phone on the table can bridge it.
Try it
Take the prototype for a spin
It's early and imperfect — that's the point. Push on it, then tell us where it breaks.
Try the live translator
English ⇄ Haitian Creole · speak, and it answers aloud in the other language
Pick two languages, press start, and put your phone on the table between you. Speak normally — when you stop, it speaks your words aloud in the other language. Then the other person replies, and it translates back. This is an early prototype, so expect a beat of delay and the occasional stumble.
Sign in to try itUses your microphone. Sessions last up to 5minutes. Best with one device and headphones off, in a quiet spot. Don't share passwords or private details.
The roadblocks
The honest hard parts
These are the obstacles between a neat demo and something you'd trust for real — and how we're attacking each one.
The echo loop
One shared speaker and one mic means the device hears its own translated voice and tries to translate that too — an endless feedback loop.
Acoustic echo cancellation plus half-duplex gating: the mic mutes itself while the device is speaking, then reopens.
Turn-taking with no button
With no push-to-talk, how does it know when one person finished and the other started — across pauses, crosstalk, and overlapping speech?
Server-side voice-activity detection from the Realtime API, tuned for the give-and-take of a real conversation.
Which direction, right now?
It has to detect which of the two languages is being spoken this instant — including Spanglish, code-switching, and similar-sounding languages.
Lock the two-language pair up front, then let the model auto-detect and translate into the other one.
Latency
Translation has to land in about a second, or people talk over it and the conversation falls apart.
Speech-to-speech streaming over WebRTC instead of a slow record → transcribe → translate → speak chain.
Barge-in and interruptions
Real conversation overlaps. People cut in, correct themselves, and change their minds mid-sentence.
The model stops speaking gracefully the moment someone jumps in, then picks up the new thread.
The words that matter most
Names, prices, addresses — and for a church, scripture terms like grace, covenant, and the names of the books of the Bible — have to come through exactly right.
A glossary and context layer that pins down the terms a generic translator tends to fumble.
Connection and cost
Live translation needs a solid network and real-time compute, which is pricey — a concern for outdoor events or mission trips.
Session caps and metering now; on-device and hybrid fallbacks under investigation for offline settings.
The tech
What we're building it with
OpenAI Realtime API
Speech-to-speech with built-in turn detection and barge-in — the core that makes sub-second, hands-free translation possible.
Gemini Live (in comparison)
Tested side-by-side as a fallback, since different engines are stronger on different language pairs. We pick the best per pair.
WebRTC + echo cancellation
A direct browser-to-model audio link with acoustic echo cancellation and half-duplex gating to tame the open-air feedback loop.
Ephemeral, server-minted tokens
The browser never holds an API key — a dedicated AWS Lambda mints a short-lived token per session, with the persona built server-side.
What you'll learn
Tech you'll get your hands on
Helping build this isn't just watching — you work shoulder-to-shoulder with the same tools professional engineers use. Here's what you'll come away knowing.
Real-time speech AI
Work directly with the OpenAI Realtime API — streaming speech-to-speech, voice-activity detection, and barge-in — the cutting edge of conversational AI.
Prompt & instruction design
Shape how the model behaves by writing and refining its system instructions, then watch your wording change the output in real time.
WebRTC & live audio
See how a browser captures a microphone, streams audio peer-to-peer, and plays a response back — the plumbing behind every video call and voice app.
Cloud functions & API security
Learn why secret keys never touch the browser, and how a small AWS Lambda mints short-lived tokens to keep them safe.
Comparing AI models
Put OpenAI and Google's Gemini Live head-to-head on real language pairs and learn to judge which tool wins where — and why.
Audio engineering basics
Get hands-on with the hard parts of live audio: echo cancellation, latency, and turn-taking in a shared, open-air space.
Building & testing a product
Practice the full loop — ship a rough version, gather honest feedback, and decide what to fix next. Real product thinking, not just code.
Help build it
This is meant to be built in the open. Join the Open-Air Real Time Translatorgroup to follow progress, try each new version first, send feedback on where it stumbles, and — if you tinker with tech — help shape what comes next. You don't need to be an engineer; honest testing is just as valuable.
Built and tested at Deerfield Beach Church of God of Prophecy, Deerfield Beach, FL.