Parakeet Podcast Processor (P³)

Automated podcast processing with Apple Silicon optimization and local LLMs

Transform podcasts into structured summaries using cutting-edge Apple Silicon ML acceleration.

Inspired by Tomasz Tunguz's innovative podcast processing system described in his "How I AI" interview. This implementation builds on his pioneering work in automated podcast analysis for venture capital and business intelligence.

⚡ Key Features

• 🎧 Smart Audio Processing: RSS feed monitoring + ffmpeg normalization
• 🚀 Lightning Fast Transcription: Parakeet MLX (30x faster than Whisper on Apple Silicon)
• 🧠 Local LLM Analysis: Ollama integration for structured summarization
• ✍️ AI Blog Generation: Iterative writing with AP English teacher grading system
• 📱 Social Media Posts: Auto-generate Twitter and LinkedIn content
• 💾 Efficient Storage: DuckDB for fast queries and analysis
• 📊 Rich Outputs: Markdown and JSON exports with topics, themes, quotes, and company mentions
• 🔒 100% Local: No API keys required, complete privacy

🚦 Quick Start

# Prerequisites: macOS with Apple Silicon + ffmpeg + Ollama
brew install ffmpeg
# Install Ollama from https://ollama.com, then: ollama pull llama3.2

# Setup P³
python3 -m venv venv && source venv/bin/activate
pip install -e .
p3 init

# Configure feeds in config/feeds.yaml
# Then run the complete pipeline:
p3 fetch && p3 transcribe && p3 digest && p3 export

# Generate blog posts from digest (Tunguz's innovation):
p3 write --topic "AI's Impact on Software Development"

# Or run the demo script:
python demo.py

⚡ Performance

• Audio Download: ~30 seconds per episode
• Parakeet Transcription: 60 minutes audio → 1 second processing
• Ollama Analysis: Full transcript → structured summary in ~10 seconds
• Total Pipeline: ~1 minute for complete podcast processing

🏗️ Architecture

RSS → ffmpeg → Parakeet MLX → Ollama → DuckDB → Export

Optimized Stack:

• Audio: ffmpeg normalization for consistent quality
• Transcription: Parakeet MLX (Apple Silicon optimized ASR)
• Analysis: Ollama (local Llama3.2 for structured extraction)
• Storage: DuckDB (fast analytical queries)

📊 Output Example

Generated Markdown Digest:

# Podcast Digest - 2025-08-25

## Test Podcast

### All About That Bass

**Summary:** The Roland TR-808 drum machine revolutionized hip-hop and electronic music...

**Key Topics:**
- Roland TR-808 drum machine  
- Hip-hop music evolution
- Electronic music production

**Notable Quotes:**
> "I really feel the 808 kick drum was one of the first things that started shattering the rules..."

**Companies Mentioned:**
- Roland Corporation

🛠️ Commands

• p3 init - Initialize directories and database
• p3 fetch - Download episodes from RSS feeds
• p3 transcribe - Convert audio to text with Parakeet MLX
• p3 digest - Generate structured summaries with Ollama
• p3 export - Export daily digests (markdown/JSON)
• p3 write --topic "Your Topic" - Generate blog posts with AP English grading
• p3 status - Show processing pipeline status

🔧 Configuration

Edit config/feeds.yaml to add your podcast feeds:

feeds:
  - name: "Your Podcast"
    url: "https://example.com/feed.xml"
    category: "tech"

settings:
  max_episodes_per_feed: 5
  
  # Transcription (Apple Silicon optimized)
  parakeet_enabled: true
  parakeet_model: "mlx-community/parakeet-tdt-0.6b-v2"
  
  # LLM Processing (100% Local)
  llm_provider: "ollama"
  llm_model: "llama3.2:latest"

📂 Project Structure

p3/
├── p3/                    # Core package
│   ├── database.py        # DuckDB storage layer
│   ├── downloader.py      # RSS + audio download with ffmpeg
│   ├── transcriber.py     # Parakeet MLX + Whisper fallback
│   ├── cleaner.py         # Ollama LLM analysis
│   ├── exporter.py        # Markdown/JSON generation
│   └── cli.py             # Command-line interface
├── config/feeds.yaml      # Podcast feed configuration
├── data/                  # Audio files + DuckDB database
├── exports/               # Generated digests
├── digest_YYYY-MM-DD.md   # Generated markdown digests
└── digest_YYYY-MM-DD.json # Generated JSON digests

🚀 Why P³?

Performance: Parakeet MLX delivers 30x speed improvement over Whisper on Apple Silicon

Privacy: 100% local processing - your podcast data never leaves your machine

Quality: State-of-the-art ASR + structured LLM analysis produces rich, actionable summaries

Efficiency: Process hours of podcasts in minutes with optimized pipeline

Perfect for researchers, journalists, content creators, or anyone who needs to efficiently process large volumes of podcast content.

🙏 Attribution

This implementation is inspired by and builds upon the innovative work of Tomasz Tunguz, founder of Theory Ventures, who pioneered many of these techniques for automated podcast analysis in venture capital. His "AP English teacher grading system" for iterative AI writing and multi-feed podcast processing approach formed the foundation for several features in this system.

Key innovations from Tunguz's system:

• ✍️ Blog post generation with AP English teacher grading (91/100 target)
• 🔄 Iterative writing improvement loops
• 📱 Social media post generation
• 🏢 Company/startup extraction for CRM integration
• 📊 Investment thesis generation from podcast insights

Source: Tomasz Tunguz interview on "How I AI" podcast