arabic-transcript-cleanup
Installation
SKILL.md
Arabic Transcript Cleanup
Speech-to-text on Arabic dialect makes predictable mistakes. Fix them before captions, editing, or repurposing. The general filler-cutting method is in filler-removal / transcript-polisher; this skill handles the Arabic part.
1. Get a good transcript first (if the user doesn't have one)
# OpenAI Whisper (local). large-v3 is the most accurate for Arabic; medium is faster.
whisper input.mp4 --language ar --model large-v3 --word_timestamps True --output_format all
# faster-whisper (much faster, same models)
pip install faster-whisper # then use WhisperModel("large-v3").transcribe(..., language="ar", word_timestamps=True)
Always force --language ar. Auto-detect sometimes picks Persian/Urdu for short clips.
If editor-apprentice is installed, its transcribe.py already produces word-level transcripts per speech island and flags islands where speech was heard but no words came out. Clean that output here (sections 2–3), then hand the corrected words back to the apprentice's stages.