transcript-fixer_skill

Overview

This skill corrects speech-to-text transcription errors in meeting notes, lectures, and interviews using a hybrid of dictionary rules and AI-driven fixes. It learns recurring patterns to build a personalized correction database and supports mixed-language scenarios and homophone issues. The workflow is designed for repeatable, team-shared correction knowledge and fast review with visual diffs.

How this skill works

The pipeline runs in three stages: dictionary-based substitutions, AI corrections via a GLM-compatible API, and a full combined pass that learns new patterns. Corrections are stored in a local SQLite database and high-confidence recurring fixes are promoted into the dictionary. Enhanced wrappers provide automatic API key detection, output organization, and generation of HTML word-level diffs for quick review.

When to use it

• Cleaning ASR/STT output from meetings, interviews, or lectures before distribution or analysis
• Fixing homophone errors and mixed English/Chinese segments in transcripts
• Building domain-specific terminology dictionaries for finance, medical, or technical teams
• Automating repeated corrections across a corpus of transcripts
• Collaborating with a team to share and approve correction rules

Best practices

• Always initialize the local corrections database before processing transcripts
• Save every manual correction immediately to the dictionary to accelerate learning
• Run the full three-stage pipeline for best accuracy: dictionary → AI → combined
• Use the enhanced wrapper for visual HTML diffs during interactive reviews
• Review learned suggestions after 2–5 occurrences before approving into the dictionary

Example use cases

• Post-process meeting transcripts to replace consistent ASR misrecognitions with correct company or product names
• Clean lecture transcripts that mix English technical terms with another language to ensure searchable text
• Create a shared correction database for a remote team handling customer interview transcriptions
• Batch-correct a corpus of historical transcripts using domain dictionaries and AI fallback