https://github.com/haasonsaas/parakeet-podcast-processor
šļø PĀ³: Lightning-fast podcast processing with Apple Silicon optimization and local LLMs. Parakeet MLX transcription + Ollama analysis = structured podcast summaries in minutes. 100% local, no API keys required.
https://github.com/haasonsaas/parakeet-podcast-processor
apple-silicon audio-processing local-llm mlx ollama parakeet podcast podcast-processing rss transcription
Last synced: 5 months ago
JSON representation
šļø PĀ³: Lightning-fast podcast processing with Apple Silicon optimization and local LLMs. Parakeet MLX transcription + Ollama analysis = structured podcast summaries in minutes. 100% local, no API keys required.
- Host: GitHub
- URL: https://github.com/haasonsaas/parakeet-podcast-processor
- Owner: haasonsaas
- Created: 2025-08-25T18:09:02.000Z (6 months ago)
- Default Branch: main
- Last Pushed: 2025-08-25T18:43:34.000Z (6 months ago)
- Last Synced: 2025-08-25T20:24:13.407Z (6 months ago)
- Topics: apple-silicon, audio-processing, local-llm, mlx, ollama, parakeet, podcast, podcast-processing, rss, transcription
- Language: Python
- Size: 41 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Parakeet Podcast Processor (PĀ³)
**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](https://tomtunguz.com)**'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.
## ā” Key Features
- **š§ Smart Audio Processing**: RSS feed monitoring + ffmpeg normalization
- **š Lightning Fast Transcription**: Parakeet MLX (30x faster than Whisper on Apple Silicon)
- **š§ Local LLM Analysis**: Ollama integration for structured summarization
- **āļø AI Blog Generation**: Iterative writing with AP English teacher grading system
- **š± Social Media Posts**: Auto-generate Twitter and LinkedIn content
- **š¾ Efficient Storage**: DuckDB for fast queries and analysis
- **š Rich Outputs**: Markdown and JSON exports with topics, themes, quotes, and company mentions
- **š 100% Local**: No API keys required, complete privacy
## š¦ Quick Start
```bash
# Prerequisites: macOS with Apple Silicon + ffmpeg + Ollama
brew install ffmpeg
# Install Ollama from https://ollama.com, then: ollama pull llama3.2
# Setup PĀ³
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
```
## ā” Performance
- **Audio Download**: ~30 seconds per episode
- **Parakeet Transcription**: 60 minutes audio ā 1 second processing
- **Ollama Analysis**: Full transcript ā structured summary in ~10 seconds
- **Total Pipeline**: ~1 minute for complete podcast processing
## šļø Architecture
```
RSS ā ffmpeg ā Parakeet MLX ā Ollama ā DuckDB ā 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)
## š Output Example
**Generated Markdown Digest:**
```markdown
# 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
```
## š ļø 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
## š§ Configuration
Edit `config/feeds.yaml` to add your podcast feeds:
```yaml
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"
```
## š Project Structure
```
p3/
āāā p3/ # Core package
ā āāā database.py # DuckDB storage layer
ā āāā downloader.py # RSS + audio download with ffmpeg
ā āāā transcriber.py # Parakeet MLX + Whisper fallback
ā āāā cleaner.py # Ollama LLM analysis
ā āāā exporter.py # Markdown/JSON generation
ā āāā cli.py # Command-line interface
āāā config/feeds.yaml # Podcast feed configuration
āāā data/ # Audio files + DuckDB database
āāā exports/ # Generated digests
āāā digest_YYYY-MM-DD.md # Generated markdown digests
āāā digest_YYYY-MM-DD.json # Generated JSON digests
```
## š Why PĀ³?
**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.
## š **Attribution**
This implementation is inspired by and builds upon the innovative work of **[Tomasz Tunguz](https://tomtunguz.com)**, 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:**
- āļø Blog post generation with AP English teacher grading (91/100 target)
- š Iterative writing improvement loops
- š± Social media post generation
- š¢ Company/startup extraction for CRM integration
- š Investment thesis generation from podcast insights
*Source: Tomasz Tunguz interview on "How I AI" podcast*