/mega-transcriber

MEGA Transcriber Skill (Local Whisper)

> **QUICK REFERENCE** > - **Output folder**: `output/transcripts/` > - **Transcribe script**: `.venv\Scripts\python.exe scripts/mega/transcribe.py` > - **Supported formats**: mp4, mp3, wav, m4a, webm, mkv, flac, ogg, aac > - **Prerequisites**: FFmpeg on PATH + `pip install openai-whisper torch` > - **File naming**: `YYYY-MM-DD_filename_transcript.txt` and `_transcript.json` > - **Python**: ALWAYS use `.venv\Scripts\python.exe` (never bare `python`)

Purpose

Transcribe audio and video files locally using OpenAI Whisper. Outputs plain text and timestamped JSON transcripts.

Trigger Phrases

• "transcribe this video", "transcribe audio", "convert speech to text"
• "transcribe my recording", "get transcript from video"
• `/content transcribe`

Supported Formats

mp4, mp3, wav, m4a, webm, mkv, flac, ogg, aac

Prerequisites

• FFmpeg installed and on PATH
• Python packages: `openai-whisper`, `torch` (install via `pip install openai-whisper torch`)

3-Mode Workflow

PLAN Mode

1. Check whisper is available:

   python -c "import whisper; print('OK')"

2. If not installed → tell user: `pip install openai-whisper torch` 3. Check FFmpeg: `ffmpeg -version` 4. Identify input file(s) — single file or batch from a folder

CLARIFY Mode

• Ask: Which file or folder to transcribe?
• Ask: Which Whisper model? Options:
• `tiny` — Fastest, least accurate (~1GB VRAM)
• `base` — Good balance (default) (~1GB VRAM)
• `small` — Better accuracy (~2GB VRAM)
• `medium` — High accuracy (~5GB VRAM)
• `large` — Best accuracy (~10GB VRAM)
• Ask: Language? (default: auto-detect)
• Ask: Output format? (default: both .txt and .json)

IMPLEMENT Mode

Run the transcription script:

python scripts/mega/transcribe.py --input "path/to/file.mp4" --model base --output references/transcripts/auto-transcribed/

For batch:

python scripts/mega/transcribe.py --input "path/to/folder/" --model base --output references/transcripts/auto-transcribed/ --batch

Output

• `.txt` — Plain text transcript
• `.json` — Timestamped segments with start/end times
• Default location: `references/transcripts/auto-transcribed/`
• Alternative: `output/transcripts/`

Notes

• First run downloads the model (~140MB for base, ~2.9GB for large)
• GPU (CUDA) used automatically if available, falls back to CPU
• CPU transcription is slower but works on any machine