Codex Transcript Translator

A personal local CLI that turns a copied YouTube transcript txt file into a bilingual transcript:

[2:03] All right, this is CS50.
[2:03] 좋아요, 이것이 CS50입니다.

It is built on top of the official @openai/codex-sdk, so the program uses Codex from your local machine instead of calling the OpenAI API directly from custom code.

What it does

• Accepts a local txt file path as an argument
• Parses YouTube transcript lines such as 2:032분 3초All right, this is CS50
• Supports both m:ss and h:mm:ss
• Keeps every parsed line and translates it into natural Korean
• Can run one extra natural-Korean review pass after translation
• Can reopen an existing bilingual txt and polish only the Korean lines
• Writes a final *_번역본.txt file beside the original input
• Saves progress after each chunk so long transcripts can resume

Requirements

• Node.js 18+
• A working Codex login on this machine

One-time setup

cd "C:\Users\USER\OneDrive\Documents\New project\codex-transcript-translator"
npm install
npx codex login

When the login prompt appears, sign in with ChatGPT.

Usage

npm run translate -- "C:\Users\USER\Downloads\챕터 1 Introduction.txt"

Optional flags:

• --output "C:\path\custom_output.txt"
• --model gpt-5.4
• --reasoning low
• --chunk-size 100
• --max-chars 15000
• --review-pass
• --polish-existing
• --overwrite
• --fresh

Recommended for long lecture transcripts on a ChatGPT/Codex plan:

npm run translate -- "C:\Users\USER\Downloads\lecture1.txt" --reasoning low --chunk-size 100 --max-chars 15000

If you want one extra natural-Korean cleanup pass before saving:

npm run translate -- "C:\Users\USER\Downloads\lecture1.txt" --reasoning low --chunk-size 100 --max-chars 15000 --review-pass

If you already have a bilingual transcript and only want to polish the Korean lines:

npm run translate -- "C:\Users\USER\Downloads\lecture0_번역본.txt" --polish-existing --reasoning low --chunk-size 100 --max-chars 15000

The polish-only mode writes a new *_다듬기.txt file by default.

For especially short subtitle-style lines, --chunk-size matters more than --max-chars. The original 30 / 7000 split can create too many Codex turns and burn through your weekly limit faster than necessary.

Output

If the input file is:

2:032분 3초All right, this is CS50.
2:122분 12초Harvard University's introduction to computer science.

The output will look like:

[2:03] All right, this is CS50.
[2:03] 좋아요, 이것이 CS50입니다.

[2:12] Harvard University's introduction to computer science.
[2:12] 하버드 대학교의 컴퓨터 과학 입문 강좌입니다.

Notes

• The translator keeps timestamps in the final file, but asks Codex to translate only the spoken text.
• Non-timestamp lines such as chapter headings are also preserved and translated.
• A progress file is written as *.progress.json while translation is running. If the process stops midway, rerun the same command to resume.