🎙️ AI Agents - Podcast Creator with Fully Local Setup

We increasingly hear about AI Agents — i usually say that we have entered what is being called the “Agentic Spring”, a phase where trends are shifting towards developing autonomous, specialized AI rather than all-knowing generalist assistants.

For this reason, I decided to explore the emerging frameworks that are rapidly gaining traction, testing their potential, understanding how they work and what these agentic frameworks can achieve.

I wanted to test their capabilities on a non-trivial use case. So, I decided to implement a podcast generation system that extracts content from a website and transforms it into a structured audio narrative. I took inspiration from Google’s Notebook LM, simulating a similar workflow where AI agents analyze and structure content before converting it into a narrated podcast.

For this project in particular i implemented a fully local agentic podcast creation pipeline utilizing:

• Ollama to run local LLMs.
• CrewAI framework for the AI Agent orchestration.
• Coqui TTS for audio generation.
• Streamlit for creating an interactive web UI.

The application automates the entire workflow of transforming news content from a website into a structured podcast with narration. Additionally, this project highlights how to enable interaction between two distinct crews within CrewAI, making it a versatile example for adapting to other use cases involving multiple collaborating crews.

---

🛠️ How It Works

Below i reported the overall schema of the Agentic Crews and Agents:

Step 1: Provide a News URL

The user inputs a news URL. This URL becomes the starting point for scraping and processing content.

Step 2: News Processing Crew

The News Processing Crew handles:

1. Scraping: Extracting and cleaning the main content from the website.
2. Reporting: Generating a detailed report based on the scraped content.

Step 3: Podcast Creation Crew

The Podcast Creation Crew takes the processed content and performs:

1. Script Writing: Crafting a structured podcast script in Italian (or whatever language do you prefer).
2. Audio Generation: Converting the script into an audio file (MP3) using Coqui TTS.

Step 4: Save Results

• The scraped content is saved locally for traceability.
• The podcast audio is saved as an MP3 file in the podcasts directory.

---

🧩 Components and Interactions

There are 2 Crews that interact with each other: News Processing Crew and Podcast Creation Crew:

1. 👥 News Processing Crew

This crew focuses on extracting and analyzing content from the provided URL.

Agents

• Website Content Scraper:
  • Role: Extract the main content of the specified website URL.
  • Tools: Uses the ScrapeWebsiteTool for efficient web scraping.
  • Output: Cleaned and relevant content in Italian.
• Senior Reporting Analyst:
  • Role: Analyze the scraped content and generate a detailed report.
  • Skills: Expertise in turning complex data into clear, actionable reports.

Tasks

1. Scrape Website Content:
   • Description: Extract and clean the main content from the specified website.
   • Expected Output: A full report with the main topics in Italian.
2. Content Reporting:
   • Description: Analyze the scraped content and expand each topic into a detailed section or provide key insights for multiple topics.
   • Expected Output: A detailed report in Italian with approximately six paragraphs.

2. 👥 Podcast Creation Crew

This crew converts processed content into a podcast script and generates the final audio file.

Agents

• Podcast Script Writer:
  • Role: Convert the processed content into a structured and engaging podcast script.
  • Output: A natural, dialogue-based script in Italian.
• Audio Producer:
  • Role: Generate high-quality MP3 audio from the podcast script.
  • Tools: Uses a custom defined tool that exploits Coqui TTS for local text-to-speech synthesis.
  • Output: An MP3 file of the podcast narration.

Tasks

1. Script Writing:
   • Description: Create a structured podcast script based on the processed content.
   • Expected Output: A podcast script in Italian with a single speaker.
2. Audio Generation:
   • Description: Convert the podcast script into an MP3 audio file.
   • Expected Output: A high-quality MP3 file (podcasts/podcast_{timestamp}.mp3).

---

🎛️ A UI Podcast Creator

To make this project even more accessible, I also developed a UI-based version using Streamlit, allowing users to interact with the system in a more intuitive way. Instead of running commands manually, users can simply enter the required details in a web-based interface and generate a podcast effortlessly:

Key Features of the UI Version:

✅ User-friendly Interface: A web-based UI powered by Streamlit for easy interaction.

✅ Customizable Inputs: Users can specify the local LLM model, enter a news URL, and select their preferred language.

✅ Automated Workflow: The system scrapes, processes, and generates podcasts.

✅ Instant Audio Playback & Download: The generated podcast is available for immediate playback and download directly from the UI.

---

📁 Project Structure

.
├── src/
│   ├── podcast_creator_from_url_ollama_coquiTTS/
│   │   ├── tools/
│   │   │   └── text_to_speech_tool.py  # Coqui TTS integration for audio generation
│   │   └── main.py                     # main script
|   |   └── streamlit_app.py            # UI version
│   └── ...
├── output/
│   ├── scraped_news.txt                # Saved content from the news scraping task
├── podcasts/
│   └── podcast.mp3                     # Generated podcast audio file
├── voices/                             # Contains the reference wav files for voice cloning
└── README.md                           # Project documentation

---

🚀 How to Run

1. Install FFmpeg:

   • Download FFmpeg from the official FFmpeg website or use a package manager suitable for your operating system.
   • Add FFmpeg to your system’s PATH:
     • Windows: Add the bin directory of FFmpeg to your system’s environment variables under the Path.
     • Linux/MacOS: Install FFmpeg using a package manager (e.g., apt install ffmpeg on Ubuntu or brew install ffmpeg on macOS) and ensure it’s accessible in the terminal.
   • Verify installation:

     ffmpeg -version
     

     If the command outputs the version details, FFmpeg is correctly installed and configured.

2. Make sure Ollama is running on your system and to have some local LLM downloaded.
3. Clone the repository:

   git clone https://github.com/enricollen/AIAgents
   cd crewAI/podcast_creator_from_url_ollama_coquiTTS
   

4. Create and activate a virtual environment to manage dependencies:

   python -m venv venv
   source venv/bin/activate   # On Linux/MacOS
   venv\Scripts\activate      # On Windows
   

5. Install required dependencies:

   pip install -r requirements.txt
   

6. Create a .env file in the project root path (see .env_example for reference) and configure environment variables:

   LLM_MODEL=llama3.1:8b
   WEBSITE_URL=https://example.com/news-article
   

7. Run the CLI version of the app:

   cd src/podcast_creator_from_url_ollama_coquiTTS
   python main.py
   

   or the UI version:

   cd src/podcast_creator_from_url_ollama_coquiTTS
   python streamlit_app.py
   

   🔗 GitHub Repository

   Visit the project repository here for accessing the codebase (if you enjoyed this content, please consider leaving a star ⭐).

🎧 Podcast Demo

Below there are a couple of demos of the final podcast audio, created in English and Italian respectively:

Screenshots 📸

Here is a screenshot illustrating the UI version of the app: