SCRIBE.

Local macOS speech-to-text tool powered by Whisper large-v3 and Apple Silicon. It accepts one media file or a folder, transcribes every supported file in place, cleans common Whisper hallucination tails, and writes Markdown drafts with YAML frontmatter.

Folders are mirrored under the selected output directory. For example:

Input:  ~/Course/block-1/lesson.mp4
Output: ~/Downloads/Course/block-1/lesson.md

Original media is never moved, changed, or deleted.

Requirements

• macOS on Apple Silicon
• Python 3
• FFmpeg and ffprobe
• mlx-whisper
• pywebview and pyobjc for the native window

Install the dependencies:

brew install ffmpeg python
/opt/homebrew/bin/python3 -m pip install mlx-whisper pywebview pyobjc

The model mlx-community/whisper-large-v3-mlx is approximately 2.9 GB. It downloads into the Hugging Face cache automatically on the first transcription. The first run therefore takes longer and requires an internet connection.

Launch

Browser mode:

./run.sh

Then use http://localhost:8930. In a regular browser, enter absolute local paths when prompted. Native file and folder dialogs are available in the app.

Native mode:

open SCRIBE.app

Or run the launcher directly:

/opt/homebrew/bin/python3 app.py

The output folder defaults to ~/Downloads and is remembered in config.json. Only .md files are placed in automatic output.

Check

./run.sh --check

This verifies that python3, ffmpeg, ffprobe, and mlx_whisper resolve under the same PATH used by the launchers.