LLM Transcription

LLM Transcription Skill

Universal transcription tools using LLMs with optimized prompts for each purpose.

When to Use This Skill

Apply when:

• Converting images (screenshots, documents, diagrams) to markdown
• Transcribing audio files to markdown text
• Batch processing multiple files for transcription
• Preserving document structure in markdown format

Do NOT apply when:

• Evaluating LLM performance (use @llm-evaluation)
• Simple text extraction without formatting needs
• Real-time transcription (these are batch tools)

Quick Start

1. Run SETUP.md once to install dependencies
2. Choose your script:
   • Image to Markdown: transcribe-image-to-markdown.py - Ensemble + judge + refinement
   • Audio to Markdown: transcribe-audio-to-markdown.py

Available Scripts

Image Transcription:

• transcribe-image-to-markdown.py - Ensemble transcription with judge and refinement

Audio Transcription:

• transcribe-audio-to-markdown.py - Convert audio to markdown transcript

Usage Examples

Image Transcription:

python transcribe-image-to-markdown.py --input-file doc.png --output-file doc.md --verbose
python transcribe-image-to-markdown.py --input-folder ./images --output-folder ./out --initial-candidates 3 --min-score 3.5

Audio Transcription:

python transcribe-audio-to-markdown.py --input recording.mp3 --output transcript.md
python transcribe-audio-to-markdown.py --input meeting.wav --model whisper-1 --output meeting.md

Supported Formats

Images: .jpg, .jpeg, .png, .gif, .webp Audio: .mp3, .wav, .m4a, .ogg, .flac

Key Features

• Optimized prompts for each transcription type
• Structure preservation - headings, lists, tables maintained
• Batch processing support for multiple files
• Cost tracking with token usage reports
• Multiple model support - OpenAI and Anthropic

Model Recommendations for Legal Documents

Tested on EU AI Act (144 pages, German legal text, 2024-02-04):

gpt-5-mini (recommended for legal):

• Verbatim transcription preserving exact wording
• Correct regulation numbers (e.g., "2022/2065")
• Accurate technical terminology
• No [unclear] markers - confident reading
• 0 critical errors in 144 pages

gpt-5-nano (not recommended for legal):

• Rewrites/summarizes instead of transcribing verbatim
• Wrong regulation numbers (e.g., "2022/665" instead of "2022/2065")
• Wrong terminology with legal/medical implications (e.g., "Blutanalyse" instead of "Blutdruck")
• 50+ [unclear] markers indicating reading difficulties
• Content duplication and structural issues

For legal documents: Use --model gpt-5-mini --dpi 120 to avoid factual errors that could have regulatory consequences. The 120 DPI setting provides optimal balance between image quality and processing speed for document transcription.

Configuration

API keys required in ~/.llm-keys or environment variables:

• OPENAI_API_KEY - For GPT models and Whisper
• ANTHROPIC_API_KEY - For Claude models