How to Build a GPT Subtitle Proofreader to Catch Translation Errors

Background

In the global content market, a single subtitle mistake can break immersion, damage brand credibility, or even lead to rejected deliveries.

While working on subtitling projects for major broadcasters like CJ ENM, we encountered repeated issues—minor in appearance but costly in consequence.

To prevent these, we built an internal automated translation QA engine and deployed it in real workflows. The results were immediate: better quality, faster delivery, and zero rejections.

‍

‍

Tech Stack

• LETR Proofreading Engine (Proprietary by Twigfarm)
  LETR is our in-house localization QA system, purpose-built for broadcast-level subtitling.
  It detects errors, suggests fixes, and ensures delivery compliance through the following modules:
  • Subtitle formatting checks (e.g., unmatched symbols, invalid line structures)
  • Delivery standard normalization (e.g., converting <br> tags to \n)
  • Embedded GPT-based language module for grammar and spell-checking of English subtitle
  ‍
• Python
  Used for building batch processing scripts and automation pipelines

• Windows Runtime Environment
  Deliverables and tools are deployed as .exe files to integrate with broadcaster-side workflows

‍

‍

LETR System Architecture (Overview)

[Input Subtitle Directory]
           ↓
[LETR Engine]
  ├─ Format Validator
  ├─ Symbol & Rule Checker
  └─ GPT-Powered Language Module
        ├─ Contextual analysis
        ├─ Spelling & grammar detection
        └─ Suggestion report generation
           ↓
[Final SRT + Issue Reports]

‍

‍

The Problem

Before automation, our manual QA process had recurring issues:

• Mismatched brackets or symbols ([[, %%, etc.)
• Invalid characters like ~, =, +
• Typos and grammatical issues in English subtitles
• Accidental English lines inside Korean SRTs
• <br> tags causing delivery rejections

Such issues, when caught late, led to emergency fixes, schedule delays, or complete delivery failures.

‍

‍

How We Solved It

We modularized the checker and implemented the pipeline with the following tools:

1. Format Checker & Auto Fix

11_check_n_fix_format.exe

• Detects unpaired symbols and fixes line structure issues
• Generates .bak backups and ko_issues.txt reports if Korean lines contain English

‍

2. Forbidden Symbol Scanner

12_check_symbols.exe

• Detects invalid or broadcaster-disallowed characters in SRT files

‍

3. Grammar & Spell Checker (LETR GPT Module)

13_check_spelling_gpt.exe
13_check_spelling_n_grammatical_errors_gpt.py

• Applies GPT-based proofreading to English subtitles
• Returns issue reports with suggested fixes

Subtitle Block: 234
Original: He go to the market every day.
Suggested: He goes to the market every day.

‍

4. Final Delivery File Generator

14_make_final_files.py

• Produces broadcast-ready subtitle files
• <br> to \n replacements handled by 15_replace_br_to_return.py

‍

‍

Results & Outcomes

• 80% Time Saved
  What used to take an hour now finishes in under 10 minutes
• Zero Delivery Failures
  Since implementing LETR, no files were rejected due to formatting or language errors
• Higher Team Satisfaction
  Staff can now focus on creative review rather than mechanical checks

‍

Lessons Learned & What’s Next

• GPT offers consistency, nuance, and context-awareness beyond human-level QA
• Language rules and delivery standards change over time → automated rules need updates
• We're expanding LETR to support multilingual subtitle QA and web-based review interfaces for collaborative editing

‍

Pro Tip

‍Localization QA isn’t just about perfect translations—it's about safe and compliant delivery.

A good proofreading engine should always follow this cycle:
Detection → Structured Reporting → Fix Suggestion → Final Output

‍

‍

References & Tools

• OpenAI GPT-4 API
• SRT Format Specification – Wikipedia
• Internal Tools:
  • 11_check_n_fix_format.exe — Format checker
  • 13_check_spelling_gpt.exe — GPT proofreader
  • 14_make_final_files.py — Final file generator
  • 15_replace_br_to_return.py — <br> to newline converter

The best subtitle checkers don’t just fix mistakes.
They prevent them before they reach the client.

Twigfarm’s LETR engine is built for that purpose.

‍

‍

‍