ausum - Audio Summarization

Automatically transcribe YouTube videos using local AI (FluidAudio Parakeet) and generate summaries with Claude Code.

This is a tool built for macOS.

Why?

Reading is faster than watching videos. For certain types of videos I find it faster to read a detailed summary versus watching the video at a faster speed.

Features

• Local speech-to-text using FluidAudio's Parakeet model (600M parameters, 25 European languages)
• Automatic summarization with Claude, following structured format
• Privacy-first - all transcription runs locally on your Mac
• Simple CLI - one command to get transcript + summary

Prerequisites

Install required tools:

# Package managers (one-time setup)
brew install yt-dlp ffmpeg

# Claude CLI
# Follow: https://docs.anthropic.com/claude-cli

# FluidAudio (build from source)
git clone https://github.com/FluidInference/FluidAudio.git
cd FluidAudio
swift build -c release

Set environment variable:

# Add to ~/.zshrc or ~/.bashrc
export FLUIDAUDIO_PATH=~/path/to/FluidAudio

Installation

# Clone this repo
git clone https://github.com/roybotbot/ausum.git
cd ausum

# Install with pip
pip install .

# Or with pipx (recommended)
pipx install .

Usage

# Basic usage
ausum https://www.youtube.com/watch?v=VIDEO_ID

# Note: Add quotes around URL if you're having issues. 

# Override saved directory for a single run
ausum "https://www.youtube.com/watch?v=VIDEO_ID" -d ~/my-transcripts

Output files:

• <video-title>.txt - Full transcript
• <video-title>-summary.md - Structured summary

First Run

On your first run, ausum will:

1. Ask where you'd like to save transcripts (defaults to ~/Documents if it exists)
2. Save your preference to ~/.config/ausum/config.json
3. Download the Parakeet model (~600MB) from HuggingFace

Subsequent runs use your saved directory preference. You can always override it with -d.

Model Storage

The Parakeet model (~460MB) is cached in ~/Library/Application Support/FluidAudio/Models/ and persists across ausum updates. It is NOT deleted when you reinstall ausum with pipx - the cache is managed by FluidAudio, not ausum.

If you need to free up disk space, you can manually delete the cache:

rm -rf ~/Library/Application\ Support/FluidAudio/Models/parakeet*

The model will be re-downloaded on next use.

Summary Format

Summaries follow the structure defined in transcript-summary.md:

• Major sections with short headers
• Concise bullet points of key points
• Step-by-step instructions (if applicable)
• Next steps for learning more

You can edit the prompt in ausum.py if you want some customization.

License

MIT - See LICENSE file