AI System for Checking Subtitle Accessibility

The problem

In day-to-day operations, video, time-coded subtitles and the language guide arrive through different systems and incident records, preventing the team from viewing the case on a single timeline.

As a result, source links may be lost while alerts for reading speed, synchronisation and speaker identification are prepared, the case may be referred to the wrong team, and the basis for a technical decision may be unclear.

An incomplete workflow may overlook an unclear proper name or auditory context; poorly designed automation creates a risk of incorrect subtitles impairing meaning and accessibility.

What we set out to improve

• Time taken for the draft reading-speed, synchronisation and speaker-identification alerts to be ready for specialist review
• Correct referral of records involving an unclear proper name or auditory context to the appropriate specialist queue
• Recording of suggestions corrected or rejected by the subtitle editor

How the system works

1. 01 Retrieve video, time-coded subtitle and language guide data from authorised sources
2. 02 Verify the source, timestamp and record identifier
3. 03 Add the relevant asset, user, location, vehicle or content context
4. 04 Prepare the draft reading-speed, synchronisation and speaker-identification alerts
5. 05 Compare the draft with technical rules and authority limits
6. 06 Check for exceptions: an unclear proper name or auditory context
7. 07 Have the subtitle editor approve, correct or reject the draft
8. 08 Write the input, suggestion, changes and final decision to the audit log

Methods we used

• Data extraction with source and timestamp traceability from video, time-coded subtitles and the language guide
• Structured output that restricts reading-speed, synchronisation and speaker-identification alerts to fields defined by technical rules
• An exception gate that separates an unclear proper name or auditory context from system processing
• An authorised queue that shows the subtitle editor the evidence alongside the suggestion
• Duplicate-processing controls and a reversal record to mitigate the risk of incorrect subtitles impairing meaning and accessibility

Where people stay involved

Data and security

Only fields required for the task are processed from video, time-coded subtitles and the language guide. Source-system permissions are preserved, and unnecessary personal, technical and security-sensitive data is minimised before being provided to the model. Every retrieval, suggestion, human change and target-system action is written to the audit log.

Who this suits

Frequently asked questions