Offline Whisper Transcription to Obsidian Notes

Local Transcribe — Whisper to Obsidian, Fully Offline

Mac only (Apple Silicon M1–M4) — not available for Windows or Intel Macs. The local transcription engine (mlx) is Apple-Silicon-specific; there is no Windows or Intel version of this skill.

Turn recordings into notes without uploading a single second of audio anywhere.

Lectures, meetings, screen recordings — you record them and never turn them into notes, because transcribing is tedious and every cloud transcription service means uploading private audio to someone else's server. This skill sets up a transcription pipeline that runs entirely on your own Mac.

What it does

• Whisper large-v3 on Apple Silicon (via mlx_whisper) — fast, accurate, and 100% offline once the model is downloaded once.
• Custom dictionary — feed it your own jargon (course terms, product names, colleague names) for measurably better recognition of specialist vocabulary.
• Optional Claude cleanup — strips filler, repetitions and meta-comments ("recording starts now", "see you next week") from the raw transcript. Uses your own Anthropic API key; skipped cleanly if you don't set one.
• Optional slide/diagram analysis — Claude Vision reads slides, tables and diagrams out of the video and merges them into the note.
• Obsidian export — the finished note lands directly in your vault as Markdown with YAML frontmatter (date, tag, source).
• Optional Google Drive upload — auto-feed NotebookLM.

How it works

#

The core path (recording → Whisper → note) never leaves your Mac and needs no API key at all. Claude cleanup and slide analysis are opt-in add-ons: set ANTHROPIC_API_KEY and they run automatically; leave it unset and the pipeline still completes end to end, just without that polish.

Usage

Drop a recording into the input folder and run it:

python3 transcribe.py                    # processes everything in input/
python3 transcribe.py /path/to/file.mp4  # or a single file directly

Or just double-click Transkribieren.command in Finder — no terminal needed after setup.

Where recordings come from: anything that outputs .mp4 or .mov works out of the box — OBS Studio (set the recording format to mp4, or remux the default .mkv afterwards), QuickTime screen recording, local Zoom/Teams recordings, or phone videos.

Setup

1. Run bash install.sh — an interactive wizard that checks you're on Apple Silicon, installs ffmpeg and the Python dependencies, and writes config.json for you (Obsidian vault path, target folder, tag).
2. Copy woerterbuch.example.txt to woerterbuch.txt and add your own domain terms — this is the single biggest lever for recognition accuracy on specialist vocabulary.
3. Optional: export ANTHROPIC_API_KEY for cleanup and slide analysis. Optional: set up a Google OAuth credentials.json for Drive/NotebookLM export.

What's included

local-transcribe/
├── transcribe.py               # the full pipeline
├── install.sh                  # interactive setup wizard
├── Transkribieren.command      # Finder double-click launcher
├── config.example.json         # vault path, formatting, vision toggle, Drive folder
├── woerterbuch.example.txt     # domain-dictionary template
└── README.md                   # original project documentation

Honest limitations

• Apple Silicon only. The local Whisper model runs via mlx, which is Apple-Silicon-specific — this does not run on Intel Macs or other platforms.
• First run downloads the model (~3 GB) — needs internet once, fully offline afterwards.
• Cleanup and vision are only as good as your own Anthropic usage/billing — they're optional precisely so the core tool never depends on an external account.
• Not a live-transcription tool. It processes finished recordings, not a real-time stream.

All rights reserved — no open-source license attached. Contact the seller for anything beyond personal use.