LinguaForge

한국어 README

Localhost proof of concept for live YouTube translation. LinguaForge captures audio from a Chrome tab, sends it to OpenAI Realtime Translation over WebRTC, plays translated audio, renders translated captions, and exports a Markdown transcript.

This repository is a PoC, not a production service. It is intentionally bound to loopback by default and keeps the OpenAI API key on the local server.

Useful search terms: OpenAI Realtime Translation, OpenAI Codex, WebRTC translation, YouTube live translation, Chrome tab audio capture, Korean captions, speech-to-speech translation, AI interpreter, realtime subtitles.

What It Does

• Captures Chrome tab audio with getDisplayMedia().
• Creates short-lived OpenAI Realtime Translation client secrets from a local Express server.
• Streams captured audio to gpt-realtime-translate over WebRTC.
• Plays translated audio in the browser.
• Shows translated captions.
• Provides original and translated audio volume controls.
• Supports manual stop, tab-ended stop, silence timeout, and max-session timeout.
• Exports the translated transcript as Markdown.

Stack

• Node.js ESM
• Express
• Browser WebRTC APIs
• OpenAI Realtime Translation
• Chrome tab audio capture

Built With Codex

This project was developed with OpenAI Codex as the coding agent. Public environment snapshot for this repository state:

• Agent: Codex, GPT-5-based coding agent
• Local environment: macOS 26.4.1, zsh, Asia/Seoul timezone
• Runtime checked with: Node.js v24.7.0, npm 11.5.1
• App runtime: localhost Express server, default port 4000
• Repository target: public GitHub project rocosrex/LinguaForge

Quick Start

cd yt-translate-poc
npm install
cp .env.example .env

Edit .env:

OPENAI_API_KEY=sk-...
PORT=4000
HOST=127.0.0.1

Run:

npm start

Open:

http://127.0.0.1:4000

In Chrome, click Start, select the YouTube tab, and make sure tab audio sharing is enabled.

Security Notes

• .env is ignored and must not be committed.
• The server defaults to 127.0.0.1.
• Unsafe bind hosts such as 0.0.0.0 are rejected.
• /session only accepts localhost Host/Origin values.
• The browser receives only a short-lived client secret, not the OpenAI API key.

Verification Results

Latest automated test run:

npm test
24 passed, 0 failed

Local smoke checks performed:

• Server started on http://127.0.0.1:4000.
• Static page loaded successfully.
• /session returned 200 OK with a configured API key.
• Non-local Host was rejected with 403.
• HOST=0.0.0.0 was rejected before listen.
• Live browser translation was manually confirmed during PoC testing.

Cost Note

The cost estimate is not reliable yet. During one rough manual run, the observed cost appeared to be around USD 10 for about 8 minutes, but the test was not isolated enough to publish that as a real benchmark. A controlled retest is needed before documenting expected cost per minute or per hour.

Public Data Note

Raw YouTube transcript exports are intentionally not tracked in this public repository. The tracked comparison report in test-output/ contains summarized evaluation notes rather than full source transcript dumps.

Known Limitations

• gpt-realtime-translate currently uses dynamic voice adaptation, so translated voice color may shift during a session.
• Custom glossary, fixed voice selection, and custom prompting are not part of this PoC.
• Browser tests are source-contract tests; full WebRTC behavior still needs manual Chrome testing.
• Transcript export currently captures translated text only.

License

MIT