▁▂▄█▆▃▁▂▅█▇▄▂▁▃▆█▄▂▁▁▂▅▇█▅▃▁▂▄▆█▃▁▂▁▄▇█▅▂▁▃▅█▆▂▁
└────────────────────┬──────────────────────────┘
│ your mac. 40 seconds. $0.
▼
transcript.md
Transcribes stuff on your computer. For free. Forever. That's it, that's the repo.
brew install ffmpeg && pip3 install mlx-whisper
python3 scripts/transcribe.py that-thing-you-recorded.mov --title "That Thing"| Rev | $15.00 |
| Sonix | $10.00 |
| Otter Pro | $8.33/mo, and it cuts you off at 20 hrs |
| OpenAI's API | $0.36 |
| this | nothing. run it a thousand times. |
| 2.5 min voice memo | 6 seconds |
|---|---|
| 23 min podcast | 40 seconds |
| 43 min interview | 3 minutes |
| that folder of 300 files | go to bed, it'll be done |
| one file | transcribe.py x.mov --title "X" |
|---|---|
| a whole folder | bulk_transcribe.py ~/footage --out ~/notes |
| who said what | open the repo in Claude Code, say "label the speakers" |
You get a .md with frontmatter and actual paragraphs, an .srt, timestamps, and the raw JSON. Drop it in Obsidian and walk away.
Whisper hands you a wall of text and has no clue who's talking. Turns out you don't need a fancier model for that, you just need to read it and figure out who's asking questions and who's answering them. Which is a thing an LLM is good at.
So CLAUDE.md is that job, written down: fill in the frontmatter, name the speakers, flag every proper noun it probably got wrong. A prompt living in the repo next to the code. Steal it, change the fields, it's yours.
┌─────────────────┐
your voice ─────▶│ no internet │─────▶ words
│ no account │
│ no meter │
└─────────────────┘
Apple Silicon only (that's the free GPU part). MIT.