🎙️ Podcast Autopilot

AI-powered podcast summarization pipeline — transcribe, intelligently trim, and export a condensed highlight reel from any audio file using OpenAI Whisper and GPT-4 Turbo.

---

📌 Overview

Podcast Autopilot is an end-to-end audio processing pipeline that automatically transforms long-form podcast or lecture recordings into tight, high-signal highlight clips — without touching a single audio editor. You drop in an MP3 or WAV file, set a target length (e.g., 30% of the original), and the system handles everything: transcription, intelligent segment selection, audio cropping, and final export.

The tool is designed for content creators, researchers, and developers who want to extract the most valuable moments from hours of audio — automatically and at scale.

---

✨ Features

• Automatic Speech-to-Text — Transcribes audio using OpenAI Whisper with word-level timestamps
• Timestamped Transcript Formatting — Groups words into readable lines, split on natural speech pauses (>1 second)
• AI-Driven Segment Selection — Uses GPT-4 Turbo to identify and select only the highest-value segments from the transcript
• Target Duration Control — Specify an exact percentage of the original runtime to keep (default: 30%)
• Retry & Validation Logic — Automatically re-queries the model if the selected duration is off-target, ensuring accuracy within ±10%
• Audio Cropping & Stitching — Extracts the selected segments and joins them with smooth 800ms silence padding
• Dual Export — Outputs both a high-quality WAV and a 192kbps MP3 of the final edited audio
• Django-Ready Utility Module — utils.py is structured for integration into a Django web application with settings-based API key management
• WAV → MP3 Auto-Conversion — Accepts both WAV and MP3 inputs; WAV files are automatically converted before processing

---

🗂️ Project Structure

podcast-autopilot/
│
├── main.py                  # Standalone pipeline script with retry/validation logic
├── utils.py                 # Django-integrated version of the pipeline
├── test_final_1.py          # Test script variant 1
├── test_final_2.py          # Test script variant 2
├── requirements.txt         # All Python dependencies
├── .gitignore
│
├── Bishop Varden Lecture.mp3  # Sample audio file
└── Tucker.mp3                 # Sample audio file

---

🔄 How It Works

Audio File (MP3/WAV)
        │
        ▼
  [1] Whisper Transcription
      (word-level timestamps)
        │
        ▼
  [2] Timestamp Formatting
      (pause-aware line grouping)
        │
        ▼
  [3] GPT-4 Turbo Summarization
      (select key segments → JSON)
        │
        ▼
  [4] Duration Validation & Retry
      (re-query if off target by >10%)
        │
        ▼
  [5] Audio Cropping & Stitching
      (pydub segment extraction)
        │
        ▼
  [6] Export: WAV + MP3

---

🛠️ Tech Stack

Technology	Purpose
Python 3.x	Core language
OpenAI Whisper (whisper-1)	Speech-to-text transcription with word timestamps
GPT-4 Turbo	Intelligent segment selection and summarization
pydub	Audio cropping, stitching, and format conversion
ffmpeg	Audio backend for pydub
python-dotenv	Environment variable management
Django	Web framework integration (via utils.py)

---

⚙️ Setup & Installation

Prerequisites

• Python 3.9+
• ffmpeg installed and available in your system PATH
• An OpenAI API key

1. Clone the repository

git clone https://github.com/BusraRafa/podcast-autopilot.git
cd podcast-autopilot

2. Install dependencies

pip install -r requirements.txt

3. Configure your API key

Create a .env file in the project root:

OPENAI_API_KEY=your_openai_api_key_here

---

🚀 Usage

Standalone Script (main.py)

Edit the bottom of main.py to point to your audio file:

if __name__ == "__main__":
    audio_file = "./your_podcast.mp3"
    output_folder = "./output"

    result = process_audio_pipeline(audio_file, output_folder, target_percentage=30)

Then run:

python main.py

Function Signature

process_audio_pipeline(
    audio_file_path: str,   # Path to your MP3 or WAV file
    output_folder: str,     # Directory for all output files
    target_percentage: int  # % of original duration to keep (default: 30)
)

Output Files

After running, the output_folder will contain:

File	Description
demo_transcription_formatted_output.txt	Full timestamped transcript
output.json	JSON array of selected segments with timestamps
<name>_FINAL_EDITED.wav	Final highlight reel (WAV)
<name>_FINAL_EDITED.mp3	Final highlight reel (MP3, 192kbps)

---

📦 Key Dependencies

openai==2.9.0
openai-whisper==20250625
pydub==0.25.1
ffmpeg-python==0.2.0
python-dotenv==1.2.1
torch==2.9.1

See requirements.txt for the full list.

---

🧠 Design Decisions

Why keep exact wording from the transcript? The pipeline instructs GPT to never paraphrase or modify source text — all selected segments are verbatim excerpts. This ensures the cropped audio matches the selected text exactly, making the JSON-to-audio alignment reliable.

Why retry logic? LLMs don't always produce outputs of a precise length on the first attempt. The pipeline calculates the total duration of selected segments after each response and retries with a stricter prompt if the result deviates more than 10% from the target.

Why 800ms silence padding? Short silence gaps between stitched segments make the final audio sound natural rather than abruptly cut. This value is configurable in the code.

---

🔮 Potential Extensions

• Web UI via Django for drag-and-drop audio upload
• Support for YouTube URL input (via yt-dlp)
• Chapter-aware summarization for structured podcasts
• Speaker diarization to preserve only a specific speaker
• Batch processing for entire podcast RSS feeds

---

📄 License

This project is open source. Feel free to use, modify, and build upon it.

---

Built with OpenAI Whisper + GPT-4 Turbo + pydub