subtitle-correction_skill

Overview

This skill corrects subtitle (.srt) files produced by speech recognition, focusing on preserving exact timestamps and subtitle indices. It is optimized for technical content with domain-specific terminology and supports both English and Chinese subtitles. The skill flags uncertain corrections for review and can produce a diff report for verification.

How this skill works

Before making edits, the skill always asks the user to supply key terminology (proper nouns, technical terms, domain vocabulary) and confirms the received list and likely domain. It reads the uploaded .srt, detects common speech-recognition errors (phonetic mistakes, technical term mistranscriptions, mixed-language issues, and code identifier formatting) and applies consistent corrections while never changing timestamps, numbering, or entry boundaries. After correction it validates structure and offers an HTML or terminal diff report and a corrected file with a -corrected suffix.

When to use it

• You have subtitle files from programming tutorials, AI/ML courses, or other technical content.
• Recognized text shows frequent domain-specific mistranscriptions (frameworks, APIs, code identifiers).
• You need to preserve exact subtitle timing and sequence while fixing text only.
• You want an auditable diff or HTML report to review all changes.
• You prefer corrections that follow a provided list of specialized terms.

Best practices

• Provide a comma-separated list of key terms before correction (proper nouns, frameworks, code identifiers).
• If you cannot supply terms, allow the tool to proceed but review flagged uncertain edits afterwards.
• Always review the generated diff report or HTML report for any ambiguous fixes.
• For long subtitle files, process in chunks to keep contextual consistency across sections.
• Keep a backup of the original file; corrected output will use a -corrected suffix.

Example use cases

• Fixing Chinese/English mixed subtitles for an AI lecture that mentions LangChain, OpenAI, and checkpointer.
• Converting spoken code references into proper identifiers (user_001, runtime.state) without altering timestamps.
• Proofreading subtitles from a web development tutorial where framework names are frequently mistranscribed.
• Batch-correcting long course videos while generating HTML diff reports for stakeholders to review.
• Processing subtitles when the speaker uses many brand or product names that standard ASR misrenders.